Pith. sign in

REVIEW 4 major objections 4 minor 85 references

Information-Bottleneck Driven Binary Neural Network for Change Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The first binary neural network built for change detection uses a training-only information-bottleneck auxiliary module to recover accuracy lost to 1-bit quantization.

desk verdict A sensible first BNN for change detection, but the evaluation is compromised by test-set tuning and a contradiction in the paper's own table. read the letter →

arxiv 2507.03504 v2 pith:QGNMHXGF submitted 2025-07-04 cs.CV

Kaijie Yin , Zhiyuan Zhang , Shu Kong , Tian Gao , Chengzhong Xu , Hui Kong This is my paper
classification cs.CV
keywords binaryneuralnetworkchangedetectioninformationbottleneckbinarizationauxiliarymodulemutualedgedeploymentremotesensing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces BiCD, a binary neural network whose weights and activations are pinned to 1 bit, and claims it is the first network of this kind built specifically for change detection. To counter the accuracy loss that comes with binarization, BiCD adds a training-only auxiliary module guided by the Information Bottleneck principle: the module pushes the 1-bit encoder to keep input information and to separate 'interest changes' from 'noise changes' in its features. The auxiliary module is removed after training, so inference remains a pure 1-bit network with no extra compute or memory. On the TSUNAMI subset of the PCD dataset, BiCD reaches an F1-score of 86.5% with a ResNet-18 backbone, beating other 1-bit baselines by 1.1–2.9 points and matching or exceeding several full-precision change detectors. If the approach holds, it would make accurate change detection practical on edge devices.

What carries the argument

The load-bearing object is the auxiliary objective module: a compact set of four parallel MLP branches plus a convolutional output layer that maps hidden features to an aligned space Z(θ,η) during training. It implements two approximate mutual-information terms — input reconstruction loss ||Z(θ,η)−X||₁ and change-separability losses on 'interest' and 'noise' feature components — which together realize the Information Bottleneck objective's Ψ term. The machinery also includes the 1-bit change generator, a binarized convolution that produces change features from the temporal feature pyramids, so that the separability loss operates directly on change-specific features. Because the module is training-only, it adds no latency or memory at inference; its role is to shape the encoder's representations so the binary network keeps enough information to distinguish meaningful changes.

What would settle it

Train BiCD with the auxiliary module but replace the three surrogate losses (reconstruction and separability) with unrelated losses of matched magnitude; if the F1 gain persists, the improvement does not come from the information-preservation mechanism the paper invokes. A second check is to take the hyperparameters β1=1e-3 and β2=0.08 tuned on the TSUNAMI subset and apply them without retuning to the GSV, VL CMU CD, and LEVIR-CD splits; if accuracy drops sharply there, the IB objective is not transferring as a general principle.

Watch

Extended reading notes

Core claim

BiCD's claim is that the accuracy loss from binarizing a change-detection network is not inevitable: it can be recovered during training by an auxiliary objective that follows the Information Bottleneck idea. The IB objective is written as min I(X,Z) - β1 I(Z,Y) - β2 Ψ, where Ψ combines a reconstruction term I(X,Z(θ,η)) that preserves input information and a separability term I(Z(θ,η),ΔX) that pushes features of changed regions apart from unchanged ones. Since exact mutual information is intractable, the paper replaces these terms with L2 and L1 losses: reconstruction error for I(X,Z), the standard change-detection loss for I(Z,Y), and masked L1 penalties on change-generator features for the separability term. The auxiliary module that computes these losses is discarded after training, so the deployed model keeps the full efficiency of a 1-bit network. The paper reports state-of-the-art results for BNN-based change detection on street-view and remote-sensing datasets, including 86.5% F1 on the TSUNAMI subset with a 1-bit C-3PO framework.

Load-bearing premise

The paper's derivation depends on treating ordinary L1 and L2 losses as faithful stand-ins for the mutual-information terms in the Information Bottleneck objective, with no proof that these losses track the same information.

Editorial extensions

If this is right

  • Deployed BiCD keeps 1-bit weights and activations, so on an ARM Cortex-A76 edge device the 1-bit C-3PO variant runs at 158.4 ms per 256×256 input versus 392.8 ms for the real-valued model, a 2.5× speedup with the same latency as the plain 1-bit baseline.
  • BiCD's gains over standard 1-bit baselines (1.1–2.9 F1 points depending on dataset and backbone) come without any change in the inference-time architecture, so the efficiency of binary convolutions is preserved.
  • The auxiliary module improves both the Siamese encoder, via reconstruction of the input pair, and the 1-bit change generator, via separability of interest/noise change features; the best configuration uses both placements.
  • On LEVIR-CD, BiCD reaches 89.9% F1 with 2.1M parameters and 6.6 G OPs, showing the recipe also works for aligned remote-sensing imagery.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the Information Bottleneck interpretation would be to measure the information-plane trajectory of BiCD against a plain 1-bit baseline; the paper's own motivation predicts a higher I(X,Z) during the drift phase.
  • The same training-only auxiliary recipe could transfer to other binary dense-prediction tasks (semantic segmentation, depth estimation), where binarization-induced information loss is suspected to hurt accuracy.
  • Since β1 and β2 were tuned on one test subset, an out-of-the-box test on unseen change-detection datasets with the same hyperparameters would show whether the IB objective generalizes or merely overfits to PCD.
  • Replacing the L1/L2 surrogates with a neural mutual-information estimator could tighten the connection to the IB objective; if accuracy improves or β-transfer stabilizes, that would support the paper's mechanism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes BiCD, a binarized (1-bit weights and activations) change detection network built on DR-TANet and C-3PO frameworks. Its main novelty is a training-only auxiliary module guided by an Information Bottleneck-inspired objective that adds reconstruction and separability losses to the standard change detection loss. The auxiliary module is removed at inference, so the deployed model retains the binary network's efficiency. The authors report F1 improvements over existing BNN baselines on PCD (TSUNAMI and GSV), VL CMU CD, and LEVIR-CD, with a headline F1 of 86.5% on the TSUNAMI subset of PCD, and they provide edge-device latency measurements.

Significance. If the empirical claims are trustworthy, the paper makes a useful contribution: it shows that a training-only auxiliary module can improve binary change detectors without increasing inference cost, and it provides a systematic comparison across three datasets and multiple BNN baselines. The edge-latency result in Table 5 is a concrete strength, as is the removal of the auxiliary module at inference. However, the central SOTA claim is weakened by a test-set-based hyperparameter selection procedure and by the absence of multiple runs or error bars, so the significance of the reported gains is not yet established.

major comments (4)
  1. [§4.2, Fig. 6, Table 4, and §4.1] The headline result, F1=86.5% on the TSUNAMI subset, is obtained after selecting beta1, beta2, and the auxiliary module placement using 'empirical analysis of the PCD dataset's TSUNAMI subset' (Fig. 6, Sec. 4.2). This is the same test set on which the state-of-the-art comparison in Table 1 is reported, so the comparison is not an independent evaluation. Because the TSUNAMI test split consists of only 20 original panoramic images, the F1 estimate is high-variance, and no seeds, confidence intervals, or significance tests are reported. The reported gains over ReActNet and AdaBin may therefore reflect selection bias rather than a genuine advantage. The authors should select hyperparameters on a validation split (or use nested cross-validation), report best-epoch selection on validation rather than test, and provide standard deviations over multiple seeds.
  2. [§4.1, Table 1] The text states: 'Remarkably, BiCD with 1-bit C-3PO surpasses full-precision DR-TANet while matching the real-valued DR-TANet / C-3PO performance on the TSUNAMI subset.' Table 1 contradicts this: the real-valued DR-TANet TSUNAMI F1 is 87.6, the real-valued C-3PO F1 is 88.4, and BiCD with 1-bit C-3PO is 86.5. BiCD does not surpass or match either full-precision method on TSUNAMI. If the intended comparison was on GSV, that should be stated explicitly; as written, the claim is false.
  3. [§3.4, Eqs. (12)-(14)] The derivation that L1/L2 losses implement the mutual information terms in the IB objective is not justified. Eq. (13) writes I(X,Z(theta,eta)) >= H(X) - R(X|Z(theta,eta)) and then adopts L1 reconstruction loss, but no argument is given that an L1 loss is an upper bound on conditional entropy or a faithful surrogate for mutual information. Similarly, Eq. (14) asserts without support that ||Z(theta)||_2 corresponds to I(X,Z(theta)) and that the L1 terms correspond to the remaining MI terms. Furthermore, Eq. (10) defines Psi with two MI terms, while Eq. (12) expands it into three terms including I(Z(theta,eta)_n, 0), whose meaning is unclear since mutual information with a constant is zero. The paper should either provide a rigorous bound justifying these surrogates or reposition the IB discussion as motivation rather than a derivation; otherwise the claim that the method is 'Information-Bottleneck driven' is not established and the transferability of the tuned beta values is questionable.
  4. [§4.2, Table 4] The ablation in Table 4 shows that adding the full auxiliary objective to the backbone alone gives +0.1 F1, while removing the separability term from the backbone gives +0.5 F1, implying that the separability term hurts when applied to Siamese features. Yet the final configuration includes the full objective in the backbone together with the generator placement. The interaction that makes this combination beneficial is not explained. This matters because the final module placement is selected on the same test set used for the headline result; without a validation-based explanation, the configuration appears to be chosen by test-set search.
minor comments (4)
  1. [Fig. 6] The axes in Figure 6 are labeled 'value of 1' and 'value of 2'; they should be labeled beta1 and beta2 for clarity.
  2. [§4.2, Table 4] The notation 'I(Z(theta,eta)_n, 0)' in Eq. (12) is confusing because 0 appears to be a constant rather than a random variable; please clarify what distribution is intended.
  3. [§3.4, Eq. (14)] The final objective mixes an L2 norm for ||Z(theta)||_2 and L1 norms for the auxiliary terms; the reason for this asymmetry is not stated and should be explained.
  4. [§4.1, Table 1] The table reports F1 for 'FC-Siam-di' and 'FC-Siam-co' with both real-valued and, in Table 3, different citation numbers; please unify the references and ensure the reported baseline values are consistent across tables.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported 86.5% TSUNAMI SOTA is obtained by tuning β1/β2 on the TSUNAMI test subset and selecting the best test epoch; the headline claim is therefore partially circular.

  1. fitted input called prediction [Section 4.2 (Hyper-Parameter Selection), Figure 6; Section 4.1 (Results on PCD); Abstract/Introduction SOTA claim.]
    "Figure 6. The hyper-parameters β1 and β2 in our BiCD framework are optimized through empirical analysis of the PCD dataset’s TSUNAMI subset. ... Experimental results show that β2 = 0.08 delivers the best performance, achieving 86.5% F1-score on the TSUNAMI subset of the PCD dataset with C-3PO."

    The headline number is not an independent prediction: β1 and β2 are swept on the TSUNAMI subset, and the configuration with the best F1 (β2=0.08, 86.5%) is then reported as BiCD's result in Table 1 and as a new state-of-the-art in the introduction. TSUNAMI is also the test subset on which all methods are compared, and Section 4.1 states that every method is evaluated by reporting 'the F1-score of the best epoch' on that test set. Thus the 86.5% figure is selected, not predicted, and the claimed superiority over ReActNet/AdaBin is at least partially the outcome of tuning the same numbers that are then cited as evidence. This is fitted-input-as-prediction circularity, not merely a missing significance test.

full rationale

The central empirical claim of the paper — the 86.5% TSUNAMI F1 and the assertion that this establishes a new state-of-the-art for 1-bit change detection — is not an independent prediction. Section 4.2 states that β1 and β2 are optimized on the TSUNAMI subset, and it is exactly that TSUNAMI subset where the selected configuration is then reported as BiCD's headline score in Table 1 and in the Introduction. Section 4.1 compounds this by reporting the best test-set epoch for every method, so the reported gains over ReActNet and AdaBin are at least partly manufactured by the evaluation protocol rather than implied by the IB objective. I do not find other load-bearing circularity: the mutual-information-to-L1/L2 surrogate in Eqs. 12–14 is an unsupported approximation and a correctness concern, but it is not a reduction of the conclusion to its inputs; the paper's one self-citation [74] is only a contextual pointer and is not used to justify the method; and no uniqueness theorem or prior-work ansatz is invoked to force the design. The score of 6 reflects that the headline empirical claim reduces in part to a test-set fit, while the method itself retains independent architectural content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on two tuned hyperparameters and on heuristic norm-based surrogates for mutual information. No new physical entities are introduced; the auxiliary module is an architectural component, not a postulated entity.

free parameters (3)
  • β1 = 1e-3
    Weight for the L2 regularization term replacing I(X,Z(θ)); selected by grid search on the TSUNAMI test subset (Section 4.2, Figure 6).
  • β2 = 0.08
    Weight for the auxiliary reconstruction and separability losses; selected on the TSUNAMI test subset; Figure 6 shows F1 varies with β2.
  • Auxiliary module placement = backbone + 1-bit change generator
    The Table 4 ablation chooses where to attach the auxiliary objective using TSUNAMI F1; this categorical design choice is part of the tuned method.
assumptions (4)
  • ad hoc to paper L1 and L2 losses are faithful surrogates for mutual information terms in the IB objective.
    Eqs. 11-14 replace I(X,Z), I(Z,Y), and I(Z,ΔX) with norm-based losses without establishing bounds or equalities; this is the load-bearing modeling assumption.
  • domain assumption The change detection task loss L_cd equals I(Z(θ),Y).
    Section 3.4 states the mutual information term 'corresponds to' the change detection loss; this is an informal identification rather than a derivation.
  • domain assumption Feature separability can be measured by Euclidean distance between hidden features of differently labeled input pairs.
    Eq. 6 defines separability as max/min distance without linking it to detection accuracy.
  • domain assumption Binarization reduces I(X,Z) as estimated in Figure 1.
    The information plane in Figure 1 is reported without the estimation method or confidence intervals; the motivation relies on this observation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Information-Bottleneck Driven Binary Neural Network for Change Detection." pith.science (2026). https://pith.science/paper/QGNMHXGF

@misc{pith2026250703504,
  author       = {Pith},
  title        = {Pith review of: Information-Bottleneck Driven Binary Neural Network for Change Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGNMHXGF}},
  note         = {Machine review of arXiv:2507.03504}
}
read the original abstract

In this paper, we propose Binarized Change Detection (BiCD), the first binary neural network (BNN) designed specifically for change detection. Conventional network binarization approaches, which directly quantize both weights and activations in change detection models, severely limit the network's ability to represent input data and distinguish between changed and unchanged regions. This results in significantly lower detection accuracy compared to real-valued networks. To overcome these challenges, BiCD enhances both the representational power and feature separability of BNNs, improving detection performance. Specifically, we introduce an auxiliary objective based on the Information Bottleneck (IB) principle, guiding the encoder to retain essential input information while promoting better feature discrimination. Since directly computing mutual information under the IB principle is intractable, we design a compact, learnable auxiliary module as an approximation target, leading to a simple yet effective optimization strategy that minimizes both reconstruction loss and standard change detection loss. Extensive experiments on street-view and remote sensing datasets demonstrate that BiCD establishes a new benchmark for BNN-based change detection, achieving state-of-the-art performance in this domain.

Figures

Figures reproduced from arXiv: 2507.03504 by the authors.

Figure 1
Figure 1. Change detection results on the PCD dataset using a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. BiCD employs a 1-bit C-3PO framework: Dual temporal feature pyramids from a shared 1-bit backbone are merged via a 1-bit [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. PopCount counts the number of “1”s in a binary-state [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Change detection results on GSV (left) and TSUNAMI [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Change detection results on VL CMU CD (top) and LEVIR-CD (bottom) datasets, comparing 1-bit baseline and BiCD based on the C-3PO framework. Left to right: input pair (t0 & t1), error map of 1-bit baseline output, and error map of BiCD output. 0 1e-5 1e-4 1e-3 1e-2 1e-1…
Figure 6
Figure 6. Figure 6: The hyper-parameters β1 and β2 in our BiCD framework are optimized through empirical analysis of the PCD dataset’s TSUNAMI subset. C-3PO. Based on these findings, we set β2 = 0.08. Effective of Components [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 69 canonical work pages

  1. [1]

    Street-view change detection with deconvolutional networks

    Pablo Fern ´andez Alcantarilla, Simon Stent, Germ ´an Ros, Roberto Arroyo, and Riccardo Gherardi. Street-view change detection with deconvolutional networks. Autonomous Robots, 42:1301 – 1322, 2016. 6

    2016

  2. [2]

    Edge-guided recurrent convolutional neural network for multitemporal re- mote sensing image building change detection

    Beifang Bai, Wei Fu, Ting Lu, and Shutao Li. Edge-guided recurrent convolutional neural network for multitemporal re- mote sensing image building change detection. IEEE Trans- actions on Geoscience and Remote Sensing, 60:1–13, 2021. 7

    2021

  3. [3]

    Onnx: Open neural net- work exchange

    Junjie Bai, Fang Lu, Ke Zhang, et al. Onnx: Open neural net- work exchange. https://github.com/onnx/onnx ,

  4. [4]

    W. G. C. Bandara and Vishal M. Patel. A transformer-based siamese network for change detection. IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Sympo- sium, pages 207–210, 2022. 7

    2022

  5. [5]

    W. G. C. Bandara, Nithin Gopalakrishnan Nair, and Vishal M. Patel. Ddpm-cd: Denoising diffusion probabilistic models as feature extractors for remote sensing change detection. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5250–5262, 2025. 7

    work page 2025

  6. [6]

    Decoupled greedy learning of cnns

    Eugene Belilovsky, Michael Eickenberg, and Edouard Oy- allon. Decoupled greedy learning of cnns. In International Conference on Machine Learning, pages 736–745. PMLR,

    work page

  7. [7]

    Courville

    Yoshua Bengio, Nicholas L´eonard, and Aaron C. Courville. Estimating or propagating gradients through stochastic neu- rons for conditional computation.ArXiv, abs/1308.3432, 2013. 4

    arXiv 2013

  8. [8]

    Rate Distortion Theory and Data Compression, pages 1–39

    Toby Berger. Rate Distortion Theory and Data Compression, pages 1–39. Springer Vienna, Vienna, 1975. 3

    work page 1975

Show all 85 references
  1. [9]

    Xnor-net++: Im- proved binary neural networks

    Adrian Bulat and Georgios Tzimiropoulos. Xnor-net++: Im- proved binary neural networks. ArXiv, abs/1909.13863, 2019. 4

    1909 arXiv

  2. [10]

    Ivan Butakov, A. D. Tolmachev, Sofia Malanchuk, A. M. Neopryatnaya, Alexey A. Frolov, and Kirill V . Andreev. Infor- mation bottleneck analysis of deep neural networks via lossy compression. ArXiv, abs/2305.08013, 2023. 2

    2023 arXiv

  3. [11]

    Binarized spectral compressive imaging

    Yuan-Yuan Cai, Yuxing Zheng, Jing Lin, Haoqian Wang, Xin Yuan, and Yulun Zhang. Binarized spectral compressive imaging. ArXiv, abs/2305.10299, 2023. 3

    2023 arXiv

  4. [12]

    A spatial-temporal attention- based method and a new dataset for remote sensing image change detection

    Hao Chen and Zhenwei Shi. A spatial-temporal attention- based method and a new dataset for remote sensing image change detection. Remote. Sens., 12:1662, 2020. 6

    2020

  5. [13]

    Remote sensing image change detection with transformers.IEEE Transactions on Geoscience and Remote Sensing, 60:1–14, 2021

    Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers.IEEE Transactions on Geoscience and Remote Sensing, 60:1–14, 2021. 6, 7

    2021

  6. [14]

    Dr-tanet: Dynamic receptive temporal attention network for street scene change detection

    Shuo Chen, Kailun Yang, and Rainer Stiefelhagen. Dr-tanet: Dynamic receptive temporal attention network for street scene change detection. 2021 IEEE Intelligent Vehicles Symposium (IV), pages 502–509, 2021. 1, 2, 6, 7

    2021

  7. [15]

    Binaryconnect: Training deep neural networks with binary weights during propagations

    Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Informa- tion Processing Systems, 2015. 2

    2015

  8. [16]

    Rodrigo Caye Daudt, B. L. Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. 2018 25th IEEE International Conference on Image Process- ing (ICIP), pages 4063–4067, 2018. 7

    2018

  9. [17]

    Jinsong Deng, Ke Wang, Yiming Deng, and G. J. Qi. Pca- based land-use change detection and analysis using multitem- poral and multisensor satellite data. International Journal of Remote Sensing, 29:4823 – 4838, 2008. 2

    2008

  10. [18]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recog- nition, pages 248–255, 2009. 6

    2009

  11. [19]

    Training deep architec- tures without end-to-end backpropagation: A survey on the provably optimal methods

    Shiyu Duan and Jose C Principe. Training deep architec- tures without end-to-end backpropagation: A survey on the provably optimal methods. IEEE Computational Intelligence Magazine, 17(4):39–51, 2022. 2, 5

    2022

  12. [20]

    Jianfei Feng, Xianzhi Yu, Yunfan Xiao, et al. bolt. https: //github.com/huawei-noah/bolt, 2021. Accessed: 2024-12-25. 8

    2021

  13. [21]

    A lightweight collective-attention network for change detection

    Yuchao Feng, Yanyan Shao, Honghui Xu, Jinshan Xu, and Jianwei Zheng. A lightweight collective-attention network for change detection. In Proceedings of the 31st ACM Inter- national Conference on Multimedia, pages 8195–8203, 2023. 1

    2023

  14. [22]

    Binary dad-net: Binarized driveable area detection network for autonomous driving

    Alexander Frickenstein, Manoj-Rohit Vemparala, Jakob Mayr, Naveen-Shankar Nagaraja, Christian Unger, Federico Tombari, and Walter Stechele. Binary dad-net: Binarized driveable area detection network for autonomous driving. In IEEE International Conference on Robotics and Autom...

    2020

  15. [23]

    Jerry D. Gibson. Rate Distortion Theory , pages 77–92. Springer Nature Switzerland, Cham, 2025. 3

    2025

  16. [24]

    Geoffrey G. Hazel. Object-level change detection in spectral imagery. IEEE Trans. Geosci. Remote. Sens., 39:553–561,

  17. [25]

    Diff-net: Image feature difference based high- definition map change detection for autonomous driving

    Lei He, Shengjie Jiang, Xiaoqing Liang, Ning Wang, and Shiyu Song. Diff-net: Image feature difference based high- definition map change detection for autonomous driving. 2022 International Conference on Robotics and Automation (ICRA), pages 2635–2641, 2021. 1

    2022

  18. [26]

    Learning deep representations by mutual information estimation and maximization

    R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2019. 5

    2019

  19. [27]

    From w-net to cdgan: Bitemporal change detection via deep learning techniques

    Bin Hou, Qingjie Liu, Heng Wang, and Yunhong Wang. From w-net to cdgan: Bitemporal change detection via deep learning techniques. IEEE Transactions on Geoscience and Remote Sensing, 58:1790–1802, 2020. 2

    2020

  20. [28]

    A survey on information bottleneck

    Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46:5325–5344,

  21. [29]

    Binarized neural networks

    Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El- Yaniv, and Yoshua Bengio. Binarized neural networks. In Advances in Neural Information Processing Systems, 2016. 2, 6

    2016

  22. [30]

    Fast object-level change detection for vhr images

    Chunlei Huo, Zhixin Zhou, Hanqing Lu, Chunhong Pan, and Keming Chen. Fast object-level change detection for vhr images. IEEE Geoscience and Remote Sensing Letters , 7: 118–122, 2010. 2

    2010

  23. [31]

    Mahoney, and Kurt Keutzer

    Sehoon Kim, Coleman Hooper, Amir Gholami, Zhen Dong, Xiuyu Li, Sheng Shen, Michael W. Mahoney, and Kurt Keutzer. Squeezellm: Dense-and-sparse quantization. ArXiv, abs/2306.07629, 2023. 1

    2023 arXiv

  24. [32]

    Adam: A method for stochastic optimiza- tion

    Diederik P Kingma. Adam: A method for stochastic optimiza- tion. arXiv preprint arXiv:1412.6980, 2014. 6

    2014 arXiv

  25. [33]

    Pruning vs quantization: Which is better? ArXiv, abs/2307.02973, 2023

    Andrey Kuzmin, Markus Nagel, Mart van Baalen, Arash Behboodi, and Tijmen Blankevoort. Pruning vs quantization: Which is better? ArXiv, abs/2307.02973, 2023. 1

    2023 arXiv

  26. [34]

    Hierarchical paired channel fusion network for street scene change detection

    Yinjie Lei, Duo Peng, Pingping Zhang, Qiuhong Ke, and Haifeng Li. Hierarchical paired channel fusion network for street scene change detection. IEEE Transactions on Image Processing, 30:55–67, 2020. 6

    2020

  27. [35]

    Ur- ban building change detection in sar images using combined differential image and residual u-net network

    Lu Li, Chao Wang, Hong Zhang, Bo Zhang, and Fan Wu. Ur- ban building change detection in sar images using combined differential image and residual u-net network. Remote. Sens., 11:1091, 2019. 1

    2019

  28. [36]

    Zhenglai Li, Chang Tang, Lizhe Wang, and Albert Y . Zomaya. Remote sensing change detection via temporal feature interac- tion and guided refinement.IEEE Transactions on Geoscience and Remote Sensing, 60:1–11, 2022. 6

    2022

  29. [37]

    Ziming Li, Chen Yan, Ying Sun, and Qinchuan Xin. A densely attentive refinement network for change detection based on very-high-resolution bitemporal remote sensing images.IEEE Transactions on Geoscience and Remote Sensing, 60:1–18,

  30. [38]

    Bi-real net: Enhancing the perfor- mance of 1-bit cnns with improved representational capability and advanced training algorithm

    Zechun Liu, Baoyuan Wu, Wenhan Luo, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Enhancing the perfor- mance of 1-bit cnns with improved representational capability and advanced training algorithm. In European Conference on Computer Vision (ECCV), 2018. 6

    2018

  31. [39]

    Reactnet: Towards precise binary neural network with generalized activation functions

    Zechun Liu, Zhiqiang Shen, Marios Savvides, and Kwang- Ting Cheng. Reactnet: Towards precise binary neural network with generalized activation functions. In European Confer- ence on Computer Vision (ECCV), pages 143–159, 2020. 3, 4, 6

    2020

  32. [40]

    Bayesian compression for deep learning

    Christos Louizos, Karen Ullrich, and Max Welling. Bayesian compression for deep learning. ArXiv, abs/1705.08665, 2017. 2

    2017 arXiv

  33. [41]

    Hicd: Change detection in quality-varied images via hierarchical correlation distillation

    Chao Pang, Xingxing Weng, Jiang Wu, Qiang Wang, and Gui- Song Xia. Hicd: Change detection in quality-varied images via hierarchical correlation distillation. IEEE Transactions on Geoscience and Remote Sensing, 62:1–16, 2024. 1

    2024

  34. [42]

    Paranjape, Celso M

    Jay N. Paranjape, Celso M. de Melo, and Vishal M. Patel. A mamba-based siamese network for remote sensing change de- tection. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1186–1196, 2024. 7

    2025

  35. [43]

    Dual task learning by leveraging both dense correspondence and mis-correspondence for robust change detection with imperfect matches

    Jin-Man Park, Ue-Hwan Kim, Seon-Hoon Lee, and Jong- Hwan Kim. Dual task learning by leveraging both dense correspondence and mis-correspondence for robust change detection with imperfect matches. 2022 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pag...

    2022

  36. [44]

    Sedona: Search for decoupled neural networks toward greedy block-wise learning

    Myeongjang Pyeon, Jihwan Moon, Taeyoung Hahn, and Gun- hee Kim. Sedona: Search for decoupled neural networks toward greedy block-wise learning. In International Confer- ence on Learning Representations, 2020. 5

    2020

  37. [45]

    A novel change detection method for natural disaster detection and segmentation from video sequence

    Huijiao Qiao, Xue Wan, Youchuan Wan, Shengyang Li, and Wanfeng Zhang. A novel change detection method for natural disaster detection and segmentation from video sequence. Sensors (Basel, Switzerland), 20, 2020. 1

    2020

  38. [46]

    Binary neural networks: A survey

    Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks: A survey. Pattern Recognition, 105:107281, 2020. 2

    2020

  39. [47]

    Xnor-net: Imagenet classification using binary convolutional neural networks

    Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV, pages 525–542,

  40. [48]

    Modeling by shortest data description*

    Jorma Rissanen. Modeling by shortest data description*. Autom., 14:465–471, 1978. 2

    1978

  41. [49]

    Change detection from a street image pair using cnn features and superpixel seg- mentation

    Ken Sakurada and Takayuki Okatani. Change detection from a street image pair using cnn features and superpixel seg- mentation. In Procedings of the British Machine Vision Con- ference 2015. British Machine Vision Association, 2015. 6, 7

    2015

  42. [50]

    Dense optical flow based change de- tection network robust to difference of camera viewpoints

    Ken Sakurada, Weimin Wang, Nobuo Kawaguchi, and Ryosuke Nakamura. Dense optical flow based change de- tection network robust to difference of camera viewpoints. ArXiv, abs/1712.02941, 2017. 2

    2017 arXiv

  43. [51]

    Weakly supervised silhouette-based semantic scene change detection

    Ken Sakurada, Mikiya Shibuya, and Weimin Wang. Weakly supervised silhouette-based semantic scene change detection. 2020 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 6861–6867, 2018. 2, 5, 6

    2020

  44. [52]

    Unsupervised change detection in wide-field video images under low illumination

    Baoqiang Shi, Zhen Jia, Jie Yang, and Nikola Kirilov Kasabov. Unsupervised change detection in wide-field video images under low illumination. IEEE Transactions on Circuits and Systems for Video Technology, 33:1564–1576, 2023. 1

    2023

  45. [53]

    Opening the black box of deep neural networks via information

    Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. arXiv preprint arXiv:1703.00810, 2017. 1, 2, 4

    2017 arXiv

  46. [54]

    An- drew Bagnell

    Boris Sofman, Bradford Neuman, Anthony Stentz, and J. An- drew Bagnell. Anytime online novelty and change detection for mobile robots. Journal of Field Robotics, 28, 2011. 1

    2011

  47. [55]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2015. 2

    2015

  48. [56]

    Deep learning and the information bottleneck principle

    Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In 2015 ieee information theory workshop (itw), pages 1–5. IEEE, 2015. 2, 4

    2015

  49. [57]

    The information bottleneck method

    Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 2, 3

    2000 arXiv

  50. [58]

    Glu-net: Global-local universal network for dense flow and correspon- dences

    Prune Truong, Martin Danelljan, and Radu Timofte. Glu-net: Global-local universal network for dense flow and correspon- dences. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6257–6267, 2019. 2

    2020

  51. [59]

    Adabin: Improving binary neural networks with adaptive bi- nary sets

    Zhaopeng Tu, Xinghao Chen, Pengju Ren, and Yunhe Wang. Adabin: Improving binary neural networks with adaptive bi- nary sets. In European Conference on Computer Vision, 2022. 3, 6

    2022

  52. [60]

    Soft weight-sharing for neural network compression

    Karen Ullrich, Edward Meeds, and Max Welling. Soft weight-sharing for neural network compression. ArXiv, abs/1702.04008, 2017. 2

    2017 arXiv

  53. [61]

    Balamuralidhar

    Ashley Varghese, Jayavardhana Gubbi, Akshaya Ramaswamy, and P. Balamuralidhar. Changenet: A deep learning architec- ture for visual change detection. In ECCV Workshops, 2018. 6

    2018

  54. [62]

    Changes-aware transformer: Learning generalized changes representation

    Dan Wang, Licheng Jiao, Jie Chen, Shuyuan Yang, and Fang Liu. Changes-aware transformer: Learning generalized changes representation. arXiv preprint arXiv:2309.13619 ,

    arXiv

  55. [63]

    Knowledge distillation-based lightweight change detection in high-resolution remote sens- ing imagery for on-board processing

    Guoqing Wang, Ning Zhang, Jue Wang, Wenchao Liu, Yizhuang Xie, and He Chen. Knowledge distillation-based lightweight change detection in high-resolution remote sens- ing imagery for on-board processing. IEEE Journal of Se- lected Topics in Applied Earth Observations and Remote...

    2024

  56. [64]

    How to reduce change detection to semantic segmentation

    Guo-Hua Wang, Bin-Bin Gao, and Chengjie Wang. How to reduce change detection to semantic segmentation. Pattern Recognition, 2023. 2, 6, 7

    2023

  57. [65]

    Mixcd- net: A lightweight change detection network mixing features across cnn and transformer.IEEE Transactions on Geoscience and Remote Sensing, 2024

    Linlin Wang, Junping Zhang, and Lorenzo Bruzzone. Mixcd- net: A lightweight change detection network mixing features across cnn and transformer.IEEE Transactions on Geoscience and Remote Sensing, 2024. 1

    2024

  58. [66]

    Revisiting locally supervised learning: an alternative to end- to-end training

    Yulin Wang, Zanlin Ni, Shiji Song, Le Yang, and Gao Huang. Revisiting locally supervised learning: an alternative to end- to-end training. In International Conference on Learning Representations, 2021. 5

    2021

  59. [67]

    Infopro: Locally supervised deep learning by maximizing information propagation

    Yulin Wang, Zanlin Ni, Yifan Pu, Cai Zhou, Jixuan Ying, Shiji Song, and Gao Huang. Infopro: Locally supervised deep learning by maximizing information propagation. Inter- national Journal of Computer Vision, pages 1–31, 2024. 2, 5

    2024

  60. [68]

    Bidet: An ef- ficient binarized object detector

    Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An ef- ficient binarized object detector. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2046–2055, 2020. 2, 3, 5, 6

    2020

  61. [69]

    Official pytorch implementation for the paper: ”bidet: An efficient binarized object detector”

    Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Official pytorch implementation for the paper: ”bidet: An efficient binarized object detector”. https://github.com/ ZiweiWangTHU/BiDet, 2020. Accessed: 2024-12-25. 5

    2020

  62. [70]

    Building change detection based on an edge-guided convolutional neural network combined with a transformer

    Liegang Xia, Jun Chen, Jiancheng Luo, Junxia Zhang, Dezhi Yang, and Zhanfeng Shen. Building change detection based on an edge-guided convolutional neural network combined with a transformer. Remote. Sens., 14:4524, 2022. 7

    2022

  63. [71]

    Biqiao Xin, Qianchen Mao, Bingshu Wang, Jiangbin Zheng, Yong Zhao, and C. L. Philip Chen. 10k is enough: An ultra- lightweight binarized network for infrared small-target detec- tion, 2025. 4

    2025

  64. [72]

    Learning 1-bit tiny object detector with discriminative feature refine- ment

    Sheng Xu, Mingze Wang, Yanjing Li, Mingbao Lin, Baochang Zhang, David Doermann, and Xiao Sun. Learning 1-bit tiny object detector with discriminative feature refine- ment. In Forty-first International Conference on Machine Learning, 2024. 2, 5

    2024

  65. [73]

    A lightweight patch-level change detection network based on multi-layer feature compression and sensitivity-guided network pruning

    Lihui Xue, Xueqian Wang, Zhihao Wang, Gang Li, Huina Song, and Zhaohui Song. A lightweight patch-level change detection network based on multi-layer feature compression and sensitivity-guided network pruning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1

    2024

  66. [74]

    Pathfinder for low-altitude aircraft with binary neural network

    Kaijie Yin, Tianxiao Gao, and Hui Kong. Pathfinder for low-altitude aircraft with binary neural network. ArXiv, abs/2409.08824, 2024. 3

    2024 arXiv

  67. [75]

    Si-bivit: Binarizing vision transformers with spatial interaction

    Peng Yin, Xiaosu Zhu, Jingkuan Song, Lianli Gao, and Heng Tao Shen. Si-bivit: Binarizing vision transformers with spatial interaction. In Proceedings of the 32nd ACM Inter- national Conference on Multimedia, pages 8169–8178, 2024. 4

    2024

  68. [76]

    Bidense: Binariza- tion for dense prediction

    Rui Yin, Haotong Qin, Yulun Zhang, Wenbo Li, Yong Guo, Jianjun Zhu, Cheng Wang, and Biao Jia. Bidense: Binariza- tion for dense prediction. arXiv preprint arXiv:2411.10346,

    arXiv

  69. [77]

    Dgma2-net: A difference-guided multiscale aggrega- tion attention network for remote sensing change detection

    Zilu Ying, Zijun Tan, Yikui Zhai, Xudong Jia, Wenba Li, Junying Zeng, Angelo Genovese, Vincenzo Piuri, and Fabio Scotti. Dgma2-net: A difference-guided multiscale aggrega- tion attention network for remote sensing change detection. IEEE Transactions on Geoscience and Remote Se...

    2024

  70. [78]

    On the information bottleneck problems: Models, connec- tions, applications and information theoretic views

    Abdellatif Zaidi, I˜naki Estella Aguerri, and Shlomo Shamai. On the information bottleneck problems: Models, connec- tions, applications and information theoretic views. Entropy, 22, 2020. 2

    2020

  71. [79]

    Lightweight change detection in heterogeneous remote sensing images with online all-integer pruning training

    Chengyang Zhang, Weiming Li, Gang Li, Huina Song, Zhao- hui Song, Xueqian Wang, and Antonio Plaza. Lightweight change detection in heterogeneous remote sensing images with online all-integer pruning training. arXiv preprint arXiv:2405.01920, 2024. 1

    2024 arXiv

  72. [80]

    Be your own teacher: Improve the performance of convolutional neural networks via self distillation

    Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Cheng- long Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3713–3722...

    2019

  73. [81]

    Pokebnn: A binary pursuit of lightweight accuracy

    Yichi Zhang, Zhiru Zhang, and Lukasz Lew. Pokebnn: A binary pursuit of lightweight accuracy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475–12485, 2022. 3

    2022

  74. [82]

    Loss functions for image restoration with neural networks

    Hang Zhao, Orazio Gallo, Iuri Frosio, and Jan Kautz. Loss functions for image restoration with neural networks. IEEE Transactions on computational imaging, 3(1):47–57, 2016. 5

    2016

  75. [83]

    Training quantized neural networks with a full- precision auxiliary module

    Bohan Zhuang, Lingqiao Liu, Mingkui Tan, Chunhua Shen, and Ian Reid. Training quantized neural networks with a full- precision auxiliary module. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1488–1497, 2020. 1, 2

    2020

  76. [84]

    Unsupervised change de- tection for satellite images using dual-tree complex wavelet transform

    Turgay C ¸elik and Kai-Kuang Ma. Unsupervised change de- tection for satellite images using dual-tree complex wavelet transform. IEEE Transactions on Geoscience and Remote Sensing, 48:1199–1210, 2008. 2

    2008

  77. [2019]

    Accessed: 2024-12-25. 8

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.