Pith. sign in

REVIEW 4 major objections 5 minor 38 references

A diversity-based filter lets edge models match eight-times-larger training sets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-04 15:12 UTC pith:SUO2WTQM

load-bearing objection A useful but under-specified engineering result: FF filtering on SBAD gives an 8x bandwidth reduction on WALT, yet the embedding function φ that drives the whole claim is never defined, making the headline result unreproducible as written. the 4 major comments →

arxiv 2509.20484 v1 pith:SUO2WTQM submitted 2025-09-24 cs.CV

Data-Efficient Stream-Based Active Distillation for Scalable Edge Model Deployment

classification cs.CV
keywords active distillationstream-based active learningdataset pruningfarthest firstknowledge distillationedge computingobject detectiondata efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that in stream-based active distillation, where a small edge model is trained on frames pseudo-labeled by a large server-side teacher, adding a second-stage diversity filter hugely improves data efficiency. Specifically, selecting frames by farthest-first coverage of an embedding space matches the performance of using eight times more frames, and can even let the small student outperform the big teacher at high iteration counts. The practical importance: edge cameras can retrain on far fewer transmitted images, saving bandwidth and storage while maintaining or improving detection accuracy. The paper demonstrates this on 15 traffic cameras with vehicle detection.

Core claim

The central discovery is that the simple Farthest First selection algorithm, applied after a confidence-based prefilter, produces a training subset whose coverage of the latent space is what matters for distillation quality, not raw image count. In the experiments, with a collection-to-budget ratio of 8, FF's student model matches the mAP of a model trained on the full 8x-larger set and surpasses the teacher model at 3200 iterations. This shows that active distillation can be made dramatically more communication-efficient without sacrificing accuracy, as long as the selection emphasizes diversity rather than only confidence.

What carries the argument

The key machinery is a two-stage selection pipeline: first, a confidence-based stream filter (TOP-CONFIDENCE) decides in real time which frames enter a buffer; second, a deterministic Farthest First algorithm maps each buffered frame to a latent embedding via a function phi, picks the densest frame as the first seed, then iteratively adds frames that have the minimum maximum cosine similarity to already-selected frames, until the budget is met. This produces a diverse, representative subset that is then pseudo-labeled by the teacher and used for fine-tuning. The embedding diversity is the mechanism that preserves information content at a fraction of the data.

Load-bearing premise

The load-bearing premise is that the embedding function phi captures exactly the visual variation that matters for vehicle detection; the paper never specifies or ablates this mapping, so if phi is poorly matched to the task, the diversity selection would not yield the reported gains.

What would settle it

One concrete test: replace the latent-space embedding phi with a fixed random projection or with raw pixels in the Farthest First filter, keeping everything else identical. If the student's mAP on WALT drops to the unfiltered SBAD baseline, then the diversity gain depends critically on phi being a learned vision embedding rather than any distance function.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Edge camera deployments can cut server-bound training traffic by a factor of eight without losing detection accuracy.
  • The student model, trained on FF-selected frames, can reach or exceed the accuracy of the much larger teacher model at sustained training budgets.
  • Diversity-based selection saturates quickly: exploration at gamma=2 captures most of the gain, so small buffers may be enough.
  • The filter is label-free and training-free on the client, so it fits the compute and privacy constraints of edge devices.
  • A compact embedding suffices for the diversity criterion, meaning the extra filtering stage costs little in practice.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The success of FF suggests that confirmation bias in active distillation is best countered by covering the input distribution, not just by avoiding low-confidence labels; a testable follow-up would vary the confidence threshold while holding coverage fixed.
  • Because phi is never specified exactly, the method's portability to other domains hinges on choosing an embedding that captures task-relevant variation; one could test this by swapping phi for raw pixels or a randomly initialized network and measuring the drop.
  • The FF criterion could be combined with temporal or spatial priors, such as avoiding near-duplicate frames within a short time window, to further reduce redundancy on long video streams.
  • The two-stage design separates 'what to look at' from 'what to keep,' a separation that might generalize to other stream-based learning settings beyond object detection, such as segmentation or person re-identification.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes D-SBAD, a modification of stream-based active distillation (SBAD) in which the confidence-based sampler populates a buffer and a second, pool-based filter selects the final B frames for pseudo-labeling by a teacher. Benchmarks on the WALT dataset with a YOLO11n student and YOLO11x teacher compare five filters (Farthest First, TFDP, Moderate, Least Confidence, Random) against SBAD and SBAD-γ. The headline result is that Farthest First (FF) matches the accuracy of SBAD-γ while transmitting eight times fewer images, and at 3200 iterations the student can exceed the pre-trained teacher. The paper also studies the effect of the buffer multiplier γ and budgets B.

Significance. If reproducible, the 8× reduction in transmitted frames at matched training iterations is practically valuable for edge-model deployment. The controlled iteration counts in the main comparison, the use of a public multi-camera benchmark, and the label-free nature of the filters are strengths. However, the central claim currently depends on an unspecified embedding and on statistical/control details that are not fully reported, so the result is plausible but not yet verifiable.

major comments (4)
  1. [§4.5, Eqs. (1)–(2), and §6] The Farthest First filter is defined through a mapping φ that is never specified. Eq. (1) and Eq. (2) depend entirely on φ, and §6 only mentions 'compact embeddings generated by edge-compatible ViT models such as DINOv2' without naming the architecture, checkpoint, feature layer, pooling, or preprocessing. Because FF is the only new component distinguishing D-SBAD from SBAD, the reported 8× bandwidth equivalence is not reproducible without this information. Please specify the exact embedding model and configuration, or provide code.
  2. [§5.2, Fig. 3] The analysis of the candidate-set size γ confounds exploration with training compute. All models are trained for 100 epochs, so increasing γ from 1 to 12 increases the number of training iterations by the same factor. The observed gains and diminishing returns may therefore reflect additional training steps rather than the filtering/exploration mechanism. To support the claim that 'moderate exploration' is the source of the gain, the comparison should match iteration counts across γ values, as done for SBAD-γ in §5.1.
  3. [§5.1, Fig. 2] The '95% Margin of error' is not defined. It is unclear whether the error bars/shading represent standard error, a confidence interval, variation across cameras, or variation across random seeds. Without this, the claimed equivalence between FF and SBAD-γ cannot be statistically assessed. Please state the number of runs, the aggregation procedure, and the exact definition of the error measure.
  4. [§2.3, §5.1] The paper motivates D-SBAD as balancing exploration and exploitation and explicitly cites PPAL and Entropy-Diversity as related two-stage methods. These are the most natural baselines for a confidence-plus-diversity pipeline, yet they are not included in the experiments. Adding them (or narrowing the claim to the evaluated filters) would make the contribution more persuasive.
minor comments (5)
  1. [§4.4] The hyperparameters α, w, and γ are fixed without sensitivity analysis. Reporting the source of these values and their effect on the result would be helpful.
  2. [§4.5] In Eq. (3), the SCS formula appears to be missing parentheses: should it be P_{i,j} / (2 sqrt(π A_{i,j}))? Please clarify the notation.
  3. [§5.2, Fig. 3] No error bars or confidence intervals are shown in Fig. 3, so it is unclear whether the differences across γ are statistically meaningful.
  4. [General] The manuscript says 'Random selects images with a uniform probability. We repeated the experiments with six seeds,' but it is unclear whether the six-seed repetition applies to all filters or only Random.
  5. [§2.2] The reference to PPAL in the introduction is appropriate, but the sentence 'we propose a two-stage approach that balances exploration and exploitation, similar to PPAL' should explicitly note that PPAL is not evaluated.

Circularity Check

0 steps flagged

No significant circularity: D-SBAD is an empirical comparison of selection filters on an external benchmark, with no fitted-parameter derivation or self-citation chain forcing the central claim.

full rationale

The central claim—that Farthest First matches SBAD-γ with eight times fewer images—is an experimental result on the external WALT dataset, not a quantity derived from the method's inputs. The selection rules in §4.5 (Eqs. 1–2 for FF, SCS for TFDP, median-distance for Moderate, confidence thresholds) are deterministic criteria whose outputs are then evaluated by training a student on the selected frames and measuring mAP50-95. No parameter is fitted to the test set, no equation is constructed so that its output is the reported mAP, and no result is imported from the authors' prior work as a uniqueness or equivalence theorem that would make the comparison forced. The use of SBAD [5] as a baseline is a citation to prior work by some of the same authors, but it is not load-bearing in a circular sense: SBAD is an external algorithm being extended and benchmarked, and the paper does not invoke any unverified claim from [5] to justify its own novelty. The underspecification of the embedding φ in §4.5 and the gesture toward 'DINOv2' in §6 are reproducibility concerns, not circularity: even if the reader cannot rerun the experiment without the exact checkpoint, the paper's reported gains are not equivalent by construction to its assumptions. Similarly, the undefined '95% margin of error' in Fig. 2 affects statistical interpretation but does not reveal a step where a prediction is identical to an input. Because the paper is self-contained against an external benchmark and its filtering strategies are unsupervised and evaluated independently, the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on several hand-chosen hyperparameters (alpha, w, gamma) and on domain assumptions about the representativeness of WALT, the reliability of the student's confidence, and the quality of the latent embedding used by Farthest First. No new physical or mathematical entities are introduced.

free parameters (5)
  • Confidence threshold quantile alpha = 0.1
    Used in TOP-CONFIDENCE selection (Section 4.4). Adopted from prior SBAD work [5]; not fitted here but a hand-chosen hyperparameter that controls how many candidate frames enter the buffer.
  • Warm-up size w = 720
    Number of frames used to compute the 1-alpha quantile threshold (Section 4.4). Chosen by hand.
  • Collection multiplier gamma = 8 (default; swept 1,2,4,8,12)
    Ratio of collected candidate set size |S| to budget B (Section 5). Main experiments use gamma=8; the paper reports diminishing returns for larger gamma.
  • Training epochs = 100 (with inverse adjustment for larger sets)
    Training schedule fixed to 100 epochs for B-size sets, epochs scaled by 1/gamma for SBAD-gamma to match iteration counts (Section 4.2).
  • Batch size = 16
    Fixed batch size for fine-tuning (Section 4.2).
axioms (5)
  • domain assumption The WALT 15-camera dataset is representative of edge camera deployment scenarios.
    All experiments use WALT footage; the claim that D-SBAD 'enhances the scalability of edge learning systems' generalizes from this dataset.
  • domain assumption The Student model's confidence is a reliable indicator of pseudo-label quality in the streaming setting.
    The SELECT stage retains SBAD's TOP-CONFIDENCE rule; if high confidence correlates with teacher correctness, the candidate set is informative.
  • domain assumption The latent embedding phi used for Farthest First (e.g., DINOv2 features) captures visual diversity relevant to object detection.
    Section 4.5 introduces phi without specification; Section 6 claims effectiveness of 'compact embeddings generated by edge-compatible ViT models such as DINOv2'. This is the paper's weakest assumption.
  • domain assumption Simulating the temporal collection period by setting |S|=gamma*B from the dataset reproduces real streaming behavior.
    Section 5: 'The temporal collection period T is experimentally simulated, such that the collected set size is a gamma-multiple of the image budget B.'
  • domain assumption Teacher pseudo-labels are accurate enough to supervise the student after fine-tuning.
    The distillation setup assumes YOLO11x pseudo-labels are a reliable training signal for YOLO11n on WALT.

pith-pipeline@v1.3.0-alltime-deepseek · 6635 in / 11858 out tokens · 74822 ms · 2026-08-04T15:12:37.941305+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Data-Efficient Stream-Based Active Distillation for Scalable Edge Model Deployment." pith.science (2026). https://pith.science/paper/SUO2WTQM

@misc{pith2026250920484,
  author       = {Pith},
  title        = {Pith review of: Data-Efficient Stream-Based Active Distillation for Scalable Edge Model Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SUO2WTQM}},
  note         = {Machine review of arXiv:2509.20484}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Edge camera-based systems are continuously expanding, facing ever-evolving environments that require regular model updates. In practice, complex teacher models are run on a central server to annotate data, which is then used to train smaller models tailored to the edge devices with limited computational power. This work explores how to select the most useful images for training to maximize model quality while keeping transmission costs low. Our work shows that, for a similar training load (i.e., iterations), a high-confidence stream-based strategy coupled with a diversity-based approach produces a high-quality model with minimal dataset queries.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 4 linked inside Pith

  1. [1]

    INTRODUCTION Edge camera systems are a cost-effective and privacy-preserv- ing solution for visual analytics tasks such as object detection and classification. As these networks scale to improve cov- erage and resilience [1], each added sensor introduces a new visual domain, making it infeasible to rely on a single, uni- versal Deep Neural Network (DNN). ...

  2. [2]

    Online Knowledge Distillation Knowledge Distillation (KD) encompasses techniques where aStudentmodel learns from anotherTeachermodel [7]

    RELATED WORK 2.1. Online Knowledge Distillation Knowledge Distillation (KD) encompasses techniques where aStudentmodel learns from anotherTeachermodel [7]. This approach can be used to create compact models that are based on the capabilities of a more complex model, and is especially useful when deploying on edge devices, where storage and computational c...

  3. [3]

    candidate set

    D-SBAD Our framework follows a client-server dynamic, where a lightweight, pre-trainedStudentmodelθ, deployed on an edge device (e.g., a camera), processes a video streamX. To ensure theStudentremains up-to-date, it is periodically fine- tuned on a server using a curated setTof training samples selected fromX. The server hosts a larger, general-purpose Te...

  4. [4]

    bikes, cars, motorcycles, buses, and trucks

    MATERIALS AND METHODS 4.1. Dataset We use the Watch and Learn Time-lapse (W ALT) dataset [6], which consists of footage from 15 cameras capturing vehi- cle circulation in public spaces over one to five weeks. The dataset includes various lighting, weather conditions, and viewpoints, with sample counts ranging from 5,000 to 40,000 frames per week due to va...

  5. [5]

    The first imageI ∗ ∈ Sis selected as the one lying in the densest region of the latent space: I∗ = arg max I∈S X I ′∈S ⟨ϕ(I), ϕ(I′)⟩.(1)

  6. [6]

    •Training-Free Dataset Pruning (TFDP): Initially de- signed for image segmentation, TFDP [24] computes aShape Complexity Score (SCS)for each instance de- tected in an image

    The remainingB−1images are iteratively added such that they minimize the maximum cosine sim- ilarity to the selected set: Ib = arg min I ′∈S max I∈F cosim(ϕ(I), ϕ(I′)).(2) Our implementation follows Algorithm 2. •Training-Free Dataset Pruning (TFDP): Initially de- signed for image segmentation, TFDP [24] computes aShape Complexity Score (SCS)for each inst...

  7. [7]

    The temporal collection periodTis experimentally simulated, such that the collected set size is aγ-multiple of the image budget B,i.e.,|S|=γB

    RESULTS AND DISCUSSION We evaluate filtering strategies, and the ratio between col- lected and transmitted image sets. The temporal collection periodTis experimentally simulated, such that the collected set size is aγ-multiple of the image budget B,i.e.,|S|=γB. 5.1. Impact of Filtering Strategy Fig. 2 compares the average mAP50−95 for filtering strategies...

  8. [8]

    CONCLUSION We proposed a filtering step within the SBAD framework to enhance scalability by reducing the amount of transmitted data while maintaining model performance. Experiments demonstrated that the Farthest First algorithm, optimizing la- tent space coverage, provided the best filtering performance, matching or exceeding baselines using substantially...

  9. [9]

    Sensing as a service model for smart cities supported by Internet of Things,

    Charith Perera, Arkady Zaslavsky, Peter Christen, and Dimitrios Georgakopoulos, “Sensing as a service model for smart cities supported by Internet of Things,”Trans- actions on Emerging Telecommunications Technologies, vol. 25, no. 1, pp. 81–93, 2014

  10. [10]

    Camera clustering for scalable stream-based active distillation,

    Dani Manjah, Davide Cacciarelli, Christophe De Vleeschouwer, and Beno ˆıt Macq, “Camera clustering for scalable stream-based active distillation,”Expert Systems with Applications, vol. 290, pp. 128408, 2025

  11. [11]

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning, MIT Press, 2016

  12. [12]

    Ultra- lytics YOLO,

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu, “Ultra- lytics YOLO,” 2023

  13. [13]

    Stream-based active distillation for scalable model deployment,

    Dani Manjah, Davide Cacciarelli, Baptiste Standaert, Mohamed Benkedadra, Gauthier Rotsart de Hertaing, Benoˆıt Macq, St ´ephane Galland, and Christophe De Vleeschouwer, “Stream-based active distillation for scalable model deployment,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2023, pp. 4998– 5006

  14. [14]

    Walt: Watch and learn 2d amodal rep- resentation from time-lapse imagery,

    N. Dinesh Reddy, Robert Tamburo, and Srinivasa G. Narasimhan, “Walt: Watch and learn 2d amodal rep- resentation from time-lapse imagery,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 9356–9366

  15. [15]

    Distill- ing the knowledge in a neural network,

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean, “Distill- ing the knowledge in a neural network,” 2015

  16. [16]

    Model compression and acceleration for deep neural networks: The principles, progress, and challenges,

    Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang, “Model compression and acceleration for deep neural networks: The principles, progress, and challenges,” IEEE Signal Processing Magazine, 2018

  17. [17]

    Semi-supervised learning for domain adaptation,

    Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada, “Semi-supervised learning for domain adaptation,” inProceedings of the AAAI Conference on Artificial Intelligence, 2019, vol. 33

  18. [18]

    Domain- adaptive online active learning for real-time intelligent video analytics on edge devices,

    Michele Boldo, Mirco De Marchi, Enrico Martini, Stefano Aldegheri, and Nicola Bombieri, “Domain- adaptive online active learning for real-time intelligent video analytics on edge devices,”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Sys- tems, vol. 43, no. 11, pp. 4105–4116, Nov 2024

  19. [19]

    Pseudo-labeling and confirmation bias in deep semi-supervised learn- ing,

    Eric Arazo, Diego Ortego, Paul Albert, Noel E O’Connor, and Kevin McGuinness, “Pseudo-labeling and confirmation bias in deep semi-supervised learn- ing,” in2020 International Joint Conference on Neural Networks (IJCNN). IEEE, 2020, pp. 1–8

  20. [20]

    Aid: Active distillation machine to leverage pre-trained black-box models in private data settings,

    Trong Nghia Hoang, Shenda Hong, Cao Xiao, Bryan Low, and Jimeng Sun, “Aid: Active distillation machine to leverage pre-trained black-box models in private data settings,” inProceedings of the Web Conference 2021, 2021, pp. 3569–3581

  21. [21]

    Multi-stream cellular test-time adaptation of real-time models evolving in dynamic environments,

    Beno ˆıt G ´erin, Ana ¨ıs Halin, Anthony Cioppa, Maxim Henry, Bernard Ghanem, Beno ˆıt Macq, Christophe De Vleeschouwer, and Marc Van Droogenbroeck, “Multi-stream cellular test-time adaptation of real-time models evolving in dynamic environments,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June ...

  22. [22]

    A comprehen- sive survey on test-time adaptation under distribution shifts,

    Jian Liang, Ran He, and Tieniu Tan, “A comprehen- sive survey on test-time adaptation under distribution shifts,”International Journal of Computer Vision, pp. 1–34, 2024

  23. [23]

    Plug and play active learning for object detection,

    Chenhongyi Yang, Lichao Huang, and Elliot J. Crowley, “Plug and play active learning for object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 17784–17793

  24. [24]

    Deep active learning for object detection,

    Soumya Roy, Asim Unmesh, and Vinay P Namboodiri, “Deep active learning for object detection,”29th British Machine Vision Conference(BMVC), 2018

  25. [25]

    Entropy-based active learning for object detection with progressive di- versity constraint,

    Jiaxi Wu, Jiaxin Chen, and Di Huang, “Entropy-based active learning for object detection with progressive di- versity constraint,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2022, pp. 9397–9406

  26. [26]

    Turning big data into tiny data: Constant-size coresets for k-means, pca, and projective clustering,

    Dan Feldman, Melanie Schmidt, and Christian Sohler, “Turning big data into tiny data: Constant-size coresets for k-means, pca, and projective clustering,”SIAM Jour- nal on Computing, vol. 49, no. 3, pp. 601–657, 2020

  27. [27]

    Selection via proxy: Ef- ficient data selection for deep learning,

    Cody Coleman, Christopher Yeh, Stephen Mussmann, Baharan Mirzasoleiman, Peter Bailis, Percy Liang, Jure Leskovec, and Matei Zaharia, “Selection via proxy: Ef- ficient data selection for deep learning,”CoRR, vol. abs/1906.11829, 2019

  28. [28]

    Coreset selection for object de- tection,

    Hojun Lee, Suyoung Kim, Junhoo Lee, Jaeyoung Yoo, and Nojun Kwak, “Coreset selection for object de- tection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2024, pp. 7682–7691

  29. [29]

    Identifying mislabeled data using the area under the margin ranking,

    Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kil- ian Q Weinberger, “Identifying mislabeled data using the area under the margin ranking,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, Eds. 2020, vol. 33, pp. 17044–17056, Curran Associates, Inc

  30. [30]

    Deepemd: Few-shot image classification with differentiable earth mover’s distance and structured clas- sifiers,

    Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen, “Deepemd: Few-shot image classification with differentiable earth mover’s distance and structured clas- sifiers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

  31. [31]

    Moderate coreset: A universal method of data selection for real-world data-efficient deep learn- ing,

    Xiaobo Xia, Jiale Liu, Jun Yu, Xu Shen, Bo Han, and Tongliang Liu, “Moderate coreset: A universal method of data selection for real-world data-efficient deep learn- ing,” inThe Eleventh International Conference on Learning Representations, 2022

  32. [32]

    Training-free dataset pruning for instance segmenta- tion,

    Yalun Dai, Lingao Xiao, Ivor Tsang, and Yang He, “Training-free dataset pruning for instance segmenta- tion,” inThe Thirteenth International Conference on Learning Representations, 2025

  33. [33]

    Microsoft COCO: common ob- jects in context,

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C. Lawrence Zitnick, “Microsoft COCO: common ob- jects in context,”CoRR, vol. abs/1405.0312, 2014

  34. [34]

    Deep active learn- ing over the long tail,

    Yonatan Geifman and Ran El-Yaniv, “Deep active learn- ing over the long tail,”CoRR, vol. abs/1711.00941, 2017

  35. [35]

    Confidence-based ac- tive learning,

    Mingkun Li and Ishwar K Sethi, “Confidence-based ac- tive learning,”IEEE transactions on pattern analysis and machine intelligence, vol. 28, no. 8, pp. 1251–1261, 2006

  36. [36]

    Robust image hashing based on radial vari- ance of pixels,

    C. De Roover, C. De Vleeschouwer, F. Lefebvre, and B. Macq, “Robust image hashing based on radial vari- ance of pixels,” inIEEE International Conference on Image Processing 2005, Sep. 2005, vol. 3, pp. III–77

  37. [37]

    Secure and low cost selective encryption for jpeg2000,

    Ayoub Massoudi, Fr ´ed´eric Lef `ebvre, Christophe De Vleeschouwer, and Francois-Olivier Devaux, “Secure and low cost selective encryption for jpeg2000,” in 2008 Tenth IEEE International Symposium on Multime- dia, Dec 2008, pp. 31–38

  38. [38]

    Adapting jpeg2000 bit allocation to pre- serve features of interest,

    C ´edric Verleysen, Nathalie Merlin, and Christophe De Vleeschouwer, “Adapting jpeg2000 bit allocation to pre- serve features of interest,” in2011 4th International Congress on Image and Signal Processing, Oct 2011, vol. 2, pp. 602–606