Pith. sign in

REVIEW 3 major objections 5 minor 55 references

A biologically inspired separable learning vision model for real-time traffic object perception in Dark

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a biologically inspired, separable-learning vision model beats real-time low-light traffic perception baselines with less compute, and that its new Dark-traffic dataset supports detection, segmentation, and optical flo

desk verdict Useful new synthetic dark-traffic benchmark and a solid efficient model, but the benchmark's realism is unvalidated and the FSLConv theory has an algebra error. read the letter →

arxiv 2509.05012 v1 pith:JYAGGS62 submitted 2025-09-05 cs.CV

Hulin Li , Qiliang Ren , Jun Li , Hanbing Wei , Zheng Liu , Linfang Fan This is my paper · ORCID
classification cs.CV
keywords low-lighttrafficperceptionDark-trafficdatasetobjectdetectioninstancesegmentationopticalflowestimationbio-inspiredvisionfeature-levelseparablelearningreal-time
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 tries to establish two things: that Dark-traffic, a new benchmark of 10,425 low-light images with 99,014 instance masks plus optical flow annotations, is large and realistic enough to train and test traffic perception in the dark; and that SLVM, a model inspired by pupillary dilation and the brain's separable visual pathways, can beat real-time detection and segmentation baselines while using less computation. The authors also show that adding a tiny pupil-dilation module to optical flow networks lowers endpoint error. A sympathetic reader would care because low-light driving perception is still unreliable, and conventional enhancement pipelines are too slow for real-time use.

What carries the argument

The argument runs on four mechanisms. LAPM simulates pupil dilation by amplifying pixel values, converts the result to grayscale, builds a binary photosensitive mask, and extracts texture features with a 1x1 convolution; it has only four trainable parameters and adds about 0.002 GFLOPs. FSLConv splits feature channels into two serial branches; the paper derives F(g) and M(g), the FLOPs and MACs increments as functions of split number, and concludes g=2 is the point of maximum return. SNI-r refines an existing soft-weight fusion scheme by adding a 1x1 convolution and sigmoid gate at pixel scale to align multi-scale features. The Dark-traffic pipeline samples target per-channel means and stand

What would settle it

Collect real low-light traffic videos with ground-truth detection, segmentation, and flow annotations (or paired short/long exposures from a fixed camera), fine-tune SLVM and the baselines on Dark-traffic, and measure the same metrics. If SLVM's margins over RT-DETR and YOLOv12 shrink or reverse on real captures, the degradation proxy rather than the architecture carried the result.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that low-light traffic perception does not need image enhancement; it needs illumination-adaptive feature extraction and task-separable learning. SLVM is built from a light-adaptive pupillary mechanism (LAPM), a two-way feature-splitting convolution (FSLConv), task-decoupled branches, and a spatial misalignment-aware fusion module (SNI-r). With these components, SLVM-S outperforms RT-DETR by 11.2 percentage points in box AP while using 34.27% of its computational cost, exceeds YOLOv12 by 6.1 points in mask AP, and improves APmask by over 5 points on average over YOLACT and YOLO baselines on Dark-traffic. On the real LIS benchmark, an end-to-end trained SLVM

Load-bearing premise

Dark-traffic's synthetic darkening is a faithful proxy for real low-light scenes: the channel-wise affine transform copies brightness and contrast statistics from real LIS images onto COCO traffic images, and the paper's gains on Dark-traffic are assumed to transfer to actual night-time driving and optical flow.

Editorial extensions

If this is right

  • Dark-traffic becomes a shared benchmark for detection, instance segmentation, and optical flow in low-light traffic, filling the absence of large densely annotated dark-scene data.
  • SLVM-S shows real-time models can gain more than 5 percentage points in mask AP and 11.2 points in box AP over strong baselines without any enhancement preprocessing.
  • A four-parameter pupil-dilation module transfers across tasks, improving both static recognition and dense motion estimation under low light.
  • The feature-splitting analysis gives a quantitative reason to stop at two serial branches, which can inform other efficient convolutional designs.
  • Trained without enhancement or denoising, SLVM-L still beats enhancement-based pipelines on the real LIS benchmark, suggesting enhancement is not a necessary precondition for dark-scene perception.

Reading between the lines

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

  • Editorial inference: If the Dark-traffic degradation proxy holds up against real night-time captures, the same statistics-driven transformation could be applied to other well-lit traffic segmentation datasets to obtain cheap low-light training sets without manual labeling.
  • Editorial inference: The single pupil-dilation factor in LAPM could act as a test-time illumination knob, enabling a lightweight domain-adaptation strategy controlled by estimated scene brightness.
  • Editorial inference: The FLOPs/MACs cost-benefit argument for g=2 is not specific to low light; testing FSLConv on well-lit detection would help separate architecture gains from lighting-specific gains.
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

3 major / 5 minor

Summary. The paper introduces Dark-traffic, a large synthetic low-light traffic perception dataset built by transferring per-channel RGB statistics from the real LIS low-light dataset to well-lit COCO and KITTI images, and proposes SLVM, a biologically inspired architecture with three main components: LAPM (pupillary-dilation-style luminance compensation), FSLConv (feature-level separable convolution), and SNI-r (a spatial refinement of a previously proposed fusion module). The authors report state-of-the-art results on Dark-traffic for detection, instance segmentation, and optical flow, and also report strong results on the real LIS benchmark. The dataset and code are promised to be released.

Significance. If the synthetic Dark-traffic pipeline is a faithful proxy for real low-light conditions, the paper would provide a valuable large-scale benchmark covering three tasks, together with an efficient and modular architecture that appears to improve over strong real-time baselines. The real-LIS experiments are an independent and useful point in favor of the method. Notable strengths are the breadth of ablations, the release of dataset/code, and the inclusion of both static and dynamic perception tasks. However, the two load-bearing pillars of the paper—the realism of the synthetic dataset and the theoretical justification for the FSLConv design—have serious issues that need to be addressed before the central empirical claims can be accepted.

major comments (3)
  1. [§3.3.2, Eqs. (4)–(9)] Eq. (5) is algebraically wrong: substituting C1/g and C2/g into the standard FLOPs formula gives FLOPs_g = 2 C1 C2 K_h K_w H W / g^2, not 2 (C2/g) C1 K_h K_w H W. Consequently Eq. (6) is also incorrect (the correct F(g) is 2C1C2KKHW(1/g^2 - 1)), and the claimed 'reduction rate of 1/4' at g=2 does not follow; FLOPs would be reduced by 3/4 and MACs by 1/2. The empirical ablation in Table 3 may still justify g=2, but the analytical derivation in the text is unsupported and should be corrected or explicitly replaced by an empirical justification.
  2. [§3.1, Figs. 2–4] The validation of Dark-traffic is circular. The per-channel target means and standard deviations are sampled from the LIS low-light statistics, and Eq. (1) forces those moments onto COCO-traffic; Figs. 3–4 then show agreement with LIS statistics as evidence of realism. This confirms only the fitting procedure, not that the synthetic images capture real low-light appearance. The pipeline is a global channel-wise affine transform and explicitly omits noise, while real dark scenes also differ in local contrast, uneven illumination, and motion blur. Without a transfer experiment to real low-light data (e.g., training on Dark-traffic and testing on LIS, or human/object-detection evaluation on real captures), the Dark-traffic SOTA claims remain unvalidated proxies.
  3. [§4.2.5, Table 6] The optical-flow subset is generated by applying the same static degradation transform to both frames of each KITTI pair. This is a brightness/contrast remapping and does not reproduce low-light motion blur, sensor noise, or temporally varying illumination. The reported 12.37% EPE reduction for NeuFlow2+LAPM therefore measures robustness to a global photometric transform, not to real dark-scene flow artifacts. The authors should validate LAPM on a real low-light flow dataset (e.g., the 'Optical Flow in the Dark' data of Ref. [13]) or at least frame the results as a synthetic-domain study.
minor comments (5)
  1. [§3.1, Eq. (2)] Equation (2) compares R_o(i,j,c) with R_o(i,j,c) on both sides; the second term should be R_adj(i,j,c). As written, the mask is always zero.
  2. [§3.3.2, Eq. (10)] The gated activation formula is ambiguous: it appears to place the batch-normalized term over a sigmoid denominator, which is not a standard gated linear unit. Please clarify the intended operation (e.g., Act(z) * sigmoid(z)) and use consistent notation.
  3. [§4.2.2, Table 3] In the g=c row, the AP50_mask change is written as ↓8.7, but the baseline is 51.6 and the reported value is 51.4, so the change should be ↓0.2. Please correct.
  4. [Supplementary, Tables c–d] The tables are labeled 'Tabel c' and 'Tabel d'; typo. Also, the LAPM branch has no channel dimension, so its fusion with the FSL branch is not fully specified in the main text.
  5. [§3.1] The term 'physically grounded' overstates the method: the degradation is a statistical moment-matching procedure, not a physical sensor/illumination model. The discussion in §3.1 that rejects synthetic noise injection is reasonable but should be framed as a design choice, not as a settled physical argument.

Circularity Check

1 steps flagged · score 4.0 of 10

Dark-traffic realism is validated against the same LIS low-light statistics that define its degradation transform; the construction forces the reported distributional match, so the benchmark's transfer to real night scenes is unproven. Model performance comparisons remain internally valid.

  1. self definitional [Section 3.1 (Illumination degradation transfer), Eq. (1) and Figure 4]
    "we estimate the target mean μtC and standard deviation σtC by sampling from truncated normal distributions fitted to the empirical statistics of the LIS-traffic (low-light) dataset ... IadjC = clip[0,255]( (σtC/σoC)(IC − μoC) + μtC ) (1) ... The category distribution and image-wise statistics are illustrated in Figure. 4(a) and 4(b), respectively, demonstrating strong alignment with the empirical distribution of real low-light images in Figure. 2(c)."

    The target moments μtC and σtC are drawn from LIS low-light channel statistics, and Eq. (1) applies an affine transform whose explicit purpose is to give each Dark-traffic channel those same sampled mean and standard deviation (apart from clipping and the color-ratio correction). Consequently, the Figure 4 'alignment' between Dark-traffic and real LIS low-light RGB statistics is imposed by construction rather than measured independently. It verifies only the global per-channel moments that were used as inputs, not noise, local contrast, uneven illumination, motion blur, or any other property of real night scenes. Thus the 'physically grounded' realism claim for Dark-traffic is a fitted input reported as validation; the model-vs-model numbers on Dark-traffic are internally valid but their t

full rationale

The circularity is confined to the dataset-realism validation. The paper's headline performance comparisons (SLVM vs RT-DETR, YOLO variants, etc.) are genuine empirical measurements on the released Dark-traffic splits and do not reduce to the degradation statistics; those numbers could change if the benchmark were redefined. The g=2 FSLConv choice is supported both by explicit FLOP/MAC arithmetic and by an ablation, so although the 'theory' is informal, it is not a self-citation or a definitional result. SNI-r builds on the first author's SAN [21], but the comparison is empirical and the improvement over SNI is small, so the self-citation is not load-bearing. The main issue is that Dark-traffic's central claim of matching real low-light conditions is validated by the same moments used to build the transform (Eq. 1 and Figure 4), making that validation circular. The LIS experiments provide some independent real-world signal, but SLVM-L is trained on LIS itself rather than transferred from Dark-traffic, so they do not validate the synthetic benchmark. Score 4: partial, construction-level circularity in the benchmark's realism claim, while the core model-architecture results retain independent content.

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

The central claims rest on four free parameters (pupil dilation factor, two thresholds, and data-fitted degradation statistics) and four assumptions, including the key proxy assumption that affine channel statistics from LIS produce realistic dark images. No new physical entities are introduced.

free parameters (4)
  • pupil dilation factor (lambda or alpha) = reported range [5.0, 12.0], exact value not given
    LAPM amplifies pixel intensities by this factor (Appendix B, Eq. a). It is a hand-tuned, scene-dependent scalar; no automatic adaptation rule is specified, so it is a free parameter of the method.
  • photon threshold tau for LAPM binarization = 0.02
    Fixed luminance threshold for the photosensitive binary mask (Appendix B, Eq. c); chosen empirically, not derived.
  • color consistency threshold tau in Eq. (2) = 0.5
    Threshold for the binary correction mask in the degradation pipeline, stated as empirically set.
  • truncated normal parameters for target statistics = medians and stds observed on LIS-traffic
    The degradation pipeline samples target channel means and stds from truncated normals fitted to real dark LIS images; these are data-derived and fitted to the target distribution.
assumptions (4)
  • domain assumption Channel-wise affine transfer of mean and standard deviation from LIS low-light statistics produces realistic low-light images without any noise injection.
    Section 3.1 builds the entire Dark-traffic dataset on this premise. The paper does not validate against real captures beyond the statistics used to construct the images, and it explicitly rejects noise modeling without a physical definition.
  • domain assumption Ground-truth labels from well-lit source images (COCO masks, KITTI optical flow) remain valid after the degradation transform.
    The synthesis pipeline transforms COCO and KITTI images but reuses their original annotations for detection, segmentation, and flow. This assumes degradation does not change object boundaries or motion fields.
  • ad hoc to paper Biological analogies (pupillary dilation, separable learning in dogs, dorsal and ventral pathways in primates) validly translate into specific architectural components.
    Sections 3.2 and 3.3 use these analogies to motivate LAPM, FSLConv, and task-decoupled branches. No formal mapping from biology to the proposed operations is established.
  • ad hoc to paper The FLOPs and MACs formulas in Eqs. (4)-(9) correctly describe standard and grouped convolutions, and the derived optimum at g=2 follows from them.
    Eq. (5) omits a factor of 1/g in the grouped convolution FLOPs calculation, so the derived 25% reduction at g=2 is not supported by the stated algebra. The empirical ablation, not the derivation, is what supports g=2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A biologically inspired separable learning vision model for real-time traffic object perception in Dark." pith.science (2026). https://pith.science/paper/JYAGGS62

@misc{pith2026250905012,
  author       = {Pith},
  title        = {Pith review of: A biologically inspired separable learning vision model for real-time traffic object perception in Dark},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JYAGGS62}},
  note         = {Machine review of arXiv:2509.05012}
}
read the original abstract

Fast and accurate object perception in low-light traffic scenes has attracted increasing attention. However, due to severe illumination degradation and the lack of reliable visual cues, existing perception models and methods struggle to quickly adapt to and accurately predict in low-light environments. Moreover, there is the absence of available large-scale benchmark specifically focused on low-light traffic scenes. To bridge this gap, we introduce a physically grounded illumination degradation method tailored to real-world low-light settings and construct Dark-traffic, the largest densely annotated dataset to date for low-light traffic scenes, supporting object detection, instance segmentation, and optical flow estimation. We further propose the Separable Learning Vision Model (SLVM), a biologically inspired framework designed to enhance perception under adverse lighting. SLVM integrates four key components: a light-adaptive pupillary mechanism for illumination-sensitive feature extraction, a feature-level separable learning strategy for efficient representation, task-specific decoupled branches for multi-task separable learning, and a spatial misalignment-aware fusion module for precise multi-feature alignment. Extensive experiments demonstrate that SLVM achieves state-of-the-art performance with reduced computational overhead. Notably, it outperforms RT-DETR by 11.2 percentage points in detection, YOLOv12 by 6.1 percentage points in instance segmentation, and reduces endpoint error (EPE) of baseline by 12.37% on Dark-traffic. On the LIS benchmark, the end-to-end trained SLVM surpasses Swin Transformer+EnlightenGAN and ConvNeXt-T+EnlightenGAN by an average of 11 percentage points across key metrics, and exceeds Mask RCNN (with light enhancement) by 3.1 percentage points. The Dark-traffic dataset and complete code is released at https://github.com/alanli1997/slvm.

Figures

Figures reproduced from arXiv: 2509.05012 by the authors.

Figure 1
Figure 1. Dark-traffic dataset synthesis pipeline for low-light traffic scene perception via illumination style transfer. We first curated traffic-relevant samples from the real-world LIS dataset [1], manually adding annotations for underrepresented but critical classes such as person and traffic light. To analyze illumination-dependent characteristics, we conducted RGB-channel-wise feature statistics on both well-lit and low… view at source ↗
Figure 2
Figure 2. LIS-traffic analysis: (a) RGB value distribution of well-lit images. (b) Class-wise annotation distribution for the six traffic-related categories (unbalance). (c) RGB value distribution of low-lit images. These distributions form the foundation for constructing a feature-aligned illumination degradation mapping that transfers images from well-lit (source domain) to low-lit (target domain) conditions. Specifically, … view at source ↗
Figure 3
Figure 3. COCO-traffic analysis: (a) RGB value distribution of well-lit images. (b) Class-wise annotation distribution for the six traffic-related categories (balance). (c) RGB value distribution of low-lit images. Finally, we merge the transformed COCO-traffic (low-lit) with LIS-traffic (low-lit) to construct the large￾scale Dark-traffic dataset, consisting of 10,425 low-light images and 99,014 instance-level pixel annotatio… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Computational pipeline (g) and visualization (a-e) of the LAPM: (a) Input low-light image sample. (b) Simulated pupillary dilation-based luminance-enhanced image. (c) Ground-truth well-lit counterpart. (d) Grayscale version of the (b). (e) Photoreceptive binary mask ge…
Figure 6
Figure 6. Figure 6: The SLVM framework inspired by primate visual systems. The proposed architecture consists of three biologically inspired components: the Branch of LAPM, responsible for luminance-aware feature modulation through the Light Adaptive Pupillary Mechanism; the Branch of FSL…
Figure 8
Figure 8. Figure 8: Qualitative comparison of texture features under SNI and SNI-r. Visualization of texture feature representations produced by the SNI module and the proposed SNI-r module. The SNI-r demonstrates enhanced spatial texture alignment and improved feature consistency in chal…
Figure 9
Figure 9. Figure 9: , RT-DETR exhibits the weakest performance, with a considerable number of false positive detections. In contrast, the YOLO series and SLVM produce predictions that are much closer to the ground truth. Notably, SLVM consistently yields the most precise segmentation mask…
Figure 10
Figure 10. Figure 10: Qualitative comparisons on dynamic vision tasks (optical flow estimation). Models originally designed for well￾lit conditions generate blurry motion boundaries under low light, whereas SLVM-integrated models exhibit significantly improved optical flow accuracy. 5. Dis…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 48 canonical work pages

  1. [13]

    Optical flow in the dark[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zheng Y , Zhang M, Lu F. Optical flow in the dark[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 6749-6757

    work page 2020

  2. [1]

    Instance segmentation in the dark[J]

    Chen L, Fu Y , Wei K, et al. Instance segmentation in the dark[J]. International Journal of Computer Vision, 2023, 131(8): 2198-2218

    work page 2023

  3. [2]

    A deep learning framework for neuroscience[J]

    Richards B A, Lillicrap T P, Beaudoin P, et al. A deep learning framework for neuroscience[J]. Nature neuroscience, 2019, 22(11): 1761-1770

    work page 2019

  4. [3]

    Learning task-state representations[J]

    Niv Y . Learning task-state representations[J]. Nature neuroscience, 2019, 22(10): 1544-1553

    work page 2019

  5. [4]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]

    Howard A G. Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]. arXiv preprint arXiv:1704.04861, 2017

    arXiv 2017

  6. [5]

    Feature pyramid networks for object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

    Lin T Y , Dollár P, Girshick R, et al. Feature pyramid networks for object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 2117-2125

    work page 2017

  7. [6]

    You only look once: Unified, real -time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

    Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real -time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 779- 788

    work page 2016

  8. [7]

    Yolact: Real -time instance segmentation[C]//Proceedings of the IEEE/CVF international conference on computer vision

    Bolya D, Zhou C, Xiao F, et al. Yolact: Real -time instance segmentation[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2019: 9157-9166

    work page 2019

Show all 55 references
  1. [8]

    Social learning in dogs[M]//The Social Dog

    Pongrácz P. Social learning in dogs[M]//The Social Dog. Academic Press, 2014: 249-293

    2014

  2. [9]

    How dogs learn[M]

    Burch M R, Bailey J S. How dogs learn[M]. Turner Publishing Company, 2008

    2008

  3. [10]

    Getting to know low -light images with the exclusively dark dataset[J]

    Loh Y P, Chan C S. Getting to know low -light images with the exclusively dark dataset[J]. Computer Vision and Image Understanding, 2019, 178: 30-42

    2019

  4. [11]

    LISU: Low -light indoor scene understanding with joint learning of reflectance restoration[J]

    Zhang N, Nex F, Kerle N, et al. LISU: Low -light indoor scene understanding with joint learning of reflectance restoration[J]. ISPRS journal of photogrammetry and remote sensing, 2022, 183: 470-481

    2022

  5. [12]

    Learning to see in the dark[C]//Procee dings of the IEEE conference on computer vision and pattern recognition

    Chen C, Chen Q, Xu J, et al. Learning to see in the dark[C]//Procee dings of the IEEE conference on computer vision and pattern recognition. 2018: 3291-3300

    2018

  6. [14]

    Multi-scale retinex for color image enhancement[C]//Proceedings of 3rd IEEE international conference on image processing

    Rahman Z, Jobson D J, Woodell G A. Multi-scale retinex for color image enhancement[C]//Proceedings of 3rd IEEE international conference on image processing. IEEE, 1996, 3: 1003-1006

    1996

  7. [15]

    An automated multi scale retinex with color restoration f or image enhancement[C]//2012 National Conference on Communications (NCC)

    Parthasarathy S, Sankaran P. An automated multi scale retinex with color restoration f or image enhancement[C]//2012 National Conference on Communications (NCC). IEEE, 2012: 1-5

    2012

  8. [16]

    Kindling the darkness: A practical low-light image enhancer[C]//Proceedings of the 27th ACM international conference on multimedia

    Zhang Y , Zhang J, Guo X. Kindling the darkness: A practical low-light image enhancer[C]//Proceedings of the 27th ACM international conference on multimedia. 2019: 1632-1640

    2019

  9. [17]

    Zero -reference deep curve estimation for low -light image enhancement[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Guo C, Li C, Guo J, et al. Zero -reference deep curve estimation for low -light image enhancement[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 1780-1789

    2020

  10. [18]

    Fbnet: Hardware -aware efficient convnet design via differentiable neural architecture search[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wu B, Dai X, Zhang P, et al. Fbnet: Hardware -aware efficient convnet design via differentiable neural architecture search[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019: 10734-10742

    2019

  11. [19]

    Rethinking the inception architect ure for computer vision[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

    Szegedy C, Vanhoucke V , Ioffe S, et al. Rethinking the inception architect ure for computer vision[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 2818 - 2826

    2016

  12. [20]

    Shufflenet v2: Practical guidelines for efficient cnn architecture design[C]//Proceedings of the European conference on computer vision (ECCV)

    Ma N, Zhang X, Zheng H T, et al. Shufflenet v2: Practical guidelines for efficient cnn architecture design[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 116-131

    2018

  13. [21]

    Rethinking Features -Fused-Pyramid-Neck for Object Detection[C]//European Conference on Computer Vision

    Li H. Rethinking Features -Fused-Pyramid-Neck for Object Detection[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 74-90

    2024

  14. [22]

    Enlightengan: Deep light enhancement without paired supervision[J]

    Jiang Y , Gong X, Liu D, et al. Enlightengan: Deep light enhancement without paired supervision[J]. IEEE transactions on image processing, 2021, 30: 2340-2349

    2021

  15. [23]

    Li J, Li B, Tu Z, et al. Light the night: A multi -condition diffusion framework for unpaired low -light enhancement in autonomous driving[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 15205-15215

    2024

  16. [24]

    AI models collapse when trained on recursively generated data[J]

    Shumailov I, Shumaylov Z, Zhao Y , et al. AI models collapse when trained on recursively generated data[J]. Nature, 2024, 631(8022): 755-759

    2024

  17. [25]

    Adaptative machine vision with microsecond-level accurate perception beyond human retina[J]

    Li L, Li S, Wang W, et al. Adaptative machine vision with microsecond-level accurate perception beyond human retina[J]. Nature Communications, 2024, 15(1): 6261

    2024

  18. [26]

    How long is the coast of Britain? Statistical self -similarity and fractional dimension[J]

    Mandelbrot B. How long is the coast of Britain? Statistical self -similarity and fractional dimension[J]. science, 1967, 156(3775): 636-638

    1967

  19. [27]

    Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

    He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778

    2016

  20. [28]

    An image is worth 16x16 words: Transformers for image recognition at scale[J]

    Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  21. [29]

    Very deep convolutional networks for large-scale image recognition[J]

    Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  22. [30]

    CSPNet: A new backbone that can enhance learning capability of CNN[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops

    Wang C Y , Liao H Y M, Wu Y H, et al. CSPNet: A new backbone that can enhance learning capability of CNN[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. 2020: 390-391

    2020

  23. [31]

    Object vision and spatial vision: two cortical pathways[J]

    Mishkin M, Ungerleider L G, Macko K A. Object vision and spatial vision: two cortical pathways[J]. Trends in neurosciences, 1983, 6: 414-417

    1983

  24. [32]

    A dual-stream neural network explains the functional segregation of dorsal and ventral visual pathways in human brains [J]

    Choi M, Han K, Wang X, et al. A dual-stream neural network explains the functional segregation of dorsal and ventral visual pathways in human brains [J]. Advances in Neural Information Processing Systems, 2023, 36: 50408-50428

    2023

  25. [33]

    Rethinking classification and localization for object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wu Y , Chen Y , Yuan L, et al. Rethinking classification and localization for object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 10186-10195

    2020

  26. [34]

    Yolox: Exceeding yolo series in 2021[J]

    Ge Z, Liu S, Wang F, et al. Yolox: Exceeding yolo series in 2021[J]. arXiv preprint arXiv:2107.08430, 2021

    2021 arXiv

  27. [35]

    Are we ready for autonomous driving? the kitti vision benchmark suite[C]//2012 IEEE conference on computer vision and pattern recognition

    Geiger A, Lenz P, Urtasun R. Are we ready for autonomous driving? the kitti vision benchmark suite[C]//2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012: 3354-3361

    2012

  28. [36]

    Microsoft coco: Common objects in context[C]//Computer vision– ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v

    Lin T Y , Maire M, Belongie S, et al. Microsoft coco: Common objects in context[C]//Computer vision– ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v

    2014

  29. [37]

    Springer International Publishing, 2014: 740-755

    2014

  30. [38]

    Slim-neck by GSConv: A lightweight-design for real-time detector architectures[J]

    Li H, Li J, Wei H, et al. Slim-neck by GSConv: A lightweight-design for real-time detector architectures[J]. Journal of Real-Time Image Processing, 2024, 21(3): 62

    2024

  31. [39]

    Computer software

    Jocher G, Chaurasia A, Qiu J (2023) YOLO by Ultralytics (Version 8.0.0). Computer software. Gi tHub. Retrieved from https://github.com/ultralytics/ultralytics

    2023

  32. [40]

    Yolov9: Learning what you want to learn using programmable gradient information[C]//European conference on computer vision

    Wang C Y , Yeh I H, Mark Liao H Y . Yolov9: Learning what you want to learn using programmable gradient information[C]//European conference on computer vision. Cham: Springer Nature Switzerland, 2024: 1-21

    2024

  33. [41]

    Yolov10: Real -time end-to-end object detection[J]

    Wang A, Chen H, Liu L, et al. Yolov10: Real -time end-to-end object detection[J]. Advances in Neural Information Processing Systems, 2024, 37: 107984-108011

    2024

  34. [42]

    Yolov11: An overview of the key architectural enhancements[J]

    Khanam R, Hussain M. Yolov11: An overview of the key architectural enhancements[J]. arXiv preprint arXiv:2410.17725, 2024

    2024 arXiv

  35. [43]

    Yolov12: Attention -centric real -time object detectors[J]

    Tian Y , Ye Q, Doermann D. Yolov12: Attention -centric real -time object detectors[J]. arXiv preprint arXiv:2502.12524, 2025

    2025 arXiv

  36. [44]

    Detrs beat yolos on real -time object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhao Y , Lv W, Xu S, et al. Detrs beat yolos on real -time object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024: 16965-16974

    2024

  37. [45]

    Mask r -cnn[C]//Proceedings of the IEEE inter national conference on computer vision

    He K, Gkioxari G, Dollár P, et al. Mask r -cnn[C]//Proceedings of the IEEE inter national conference on computer vision. 2017: 2961-2969

    2017

  38. [46]

    A convnet for the 2020s[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu Z, Mao H, Wu C Y , et al. A convnet for the 2020s[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 11976-11986

    2022

  39. [47]

    Swin transfo rmer: Hierarchical vision transformer using shifted windows[C]//Proceedings of the IEEE/CVF international conference on computer vision

    Liu Z, Lin Y , Cao Y , et al. Swin transfo rmer: Hierarchical vision transformer using shifted windows[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2021: 10012 - 10022

    2021

  40. [48]

    Masked -attention mask transformer for universal image segmentation[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Cheng B, Misra I, Schwing A G, et al. Masked -attention mask transformer for universal image segmentation[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 1290-1299

    2022

  41. [49]

    Pointrend: Image segmentation as rendering[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Kirillov A, Wu Y , He K, et al. Pointrend: Image segmentation as rendering[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 9799-9808

    2020

  42. [50]

    Restoring extremely dark images in real time[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Lamba M, Mitra K. Restoring extremely dark images in real time[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021: 3487-3497

    2021

  43. [51]

    Semantic instance se gmentation for autonomous driving[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops

    De Brabandere B, Neven D, Van Gool L. Semantic instance se gmentation for autonomous driving[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2017: 7-9

    2017

  44. [52]

    Gmflow: Learning optical flow via global matching[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xu H, Zhang J, Cai J, et al. Gmflow: Learning optical flow via global matching[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 8121-8130

    2022

  45. [53]

    Vanillanet: the power of minimalism in deep learning[J]

    Chen H, Wang Y , Guo J, et al. Vanillanet: the power of minimalism in deep learning[J]. Advances in Neural Information Processing Systems, 2023, 36: 7050-7064

    2023

  46. [54]

    Neuflow v2: High-efficiency optical flow estimation on edge devices[J]

    Zhang Z, Gupta A, Jiang H, et al. Neuflow v2: High-efficiency optical flow estimation on edge devices[J]. arXiv preprint arXiv:2408.10161, 2024

    2024 arXiv

  47. [55]

    Accurate leukocyte detection based on deformable -DETR and multi - level feature fusion for aiding diagnosis of blood diseases[J]

    Chen Y , Zhang C, Chen B, et al. Accurate leukocyte detection based on deformable -DETR and multi - level feature fusion for aiding diagnosis of blood diseases[J]. Computers in biology and medicine, 2024, 170: 107917. Supplementary Materials Appendix A. Training hyperparameter...

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.