Differentiable Proxy Learning for Adaptive Quantization Control in H.264 Video Coding

Ivan V. Baji\'c; Qihan Xu

A learned differentiable proxy for H.264 lets adaptive quantization improve rate-task trade-offs without changing the codec.

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 →

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

T0 review · grok-4.5

2026-07-14 11:23 UTC pith:YG5DCLJN

load-bearing objection Solid engineering paper that makes H.264 QP differentiable via soft-indexed DCVC-FM proxies and shows real-codec BD-rate gains; the transfer premise is load-bearing but not fatally unmeasured. the 4 major comments →

arxiv 2607.10478 v1 pith:YG5DCLJN submitted 2026-07-11 eess.IV cs.CV

Differentiable Proxy Learning for Adaptive Quantization Control in H.264 Video Coding

Qihan Xu , Ivan V. Baji\'c This is my paper

classification eess.IV cs.CV

keywords H.264differentiable proxyadaptive quantizationsoft indexingrate-distortionmachine visionMS-SSIMsemantic segmentation

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.

Standard codecs like H.264 are non-differentiable, so gradient-based optimization of their quantization parameter for a target task is normally impossible. This paper builds a neural proxy that approximates H.264 intra coding under both a single image-wide QP and a macroblock-level QP map, making the mapping from QP to rate and reconstruction differentiable via soft indexing. Once the proxy is trained and frozen, small networks can be trained through it to choose QPs that favor perceptual quality or machine-vision accuracy; the resulting QP assignments are then applied to the real H.264 encoder. The authors show that the proxy tracks H.264 rate-distortion curves reasonably well and that the adaptive-QP policies improve rate-task performance over fixed-QP baselines, with BD-rate savings reaching about 17 percent for semantic segmentation and 15 percent for MS-SSIM. The result matters because it offers a practical way to retarget an already-deployed, widely supported codec without rewriting its core loop.

Core claim

A variable-rate learned compression model, made differentiable to the H.264 quantization parameter through soft indexing of its scaling factors and trained to match H.264 reconstructions and bit rates under both global-QP and spatial-QP regimes, can serve as a frozen surrogate that enables gradient-based training of adaptive quantization networks; those networks then improve rate-task trade-offs when their predicted QPs are executed by the real non-differentiable H.264 encoder.

What carries the argument

Soft-indexing of learnable quantization scaling factors: each continuous proxy control value is turned into a soft weighted combination of discrete scales, rendering the entire proxy differentiable with respect to the codec QP (or QP map) so that task gradients can flow back to an adaptive-quantization network.

Load-bearing premise

That gradients computed through an imperfect frozen proxy still produce QP assignments that improve the real H.264 encoder on the target task.

What would settle it

Train the same adaptive-quantization networks through the proxy, then measure BD-rate on the real H.264 encoder for segmentation or MS-SSIM; if the gains disappear or reverse relative to fixed-QP baselines, the transfer claim fails.

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

If this is right

Perceptual or task-driven bit allocation can be added to existing H.264 pipelines by training only a lightweight QP predictor through the frozen proxy.
Macroblock-level spatial QP control yields larger rate-task gains than image-level global QP control on the same tasks.
Proxy fidelity need not be perfect for the method to work; useful AQ policies can still be learned under moderate rate-distortion mismatch.
The same soft-indexing construction can be reused for other standard codecs that expose a discrete quality or QP index.
Extending the proxy to inter-frame coding would allow adaptive quantization across both space and time without altering the video codec itself.

Where Pith is reading between the lines

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

Because the proxy is trained only on intra frames, residual temporal correlation in real video may reduce the observed AQ gains once motion-compensated prediction is present.
The larger RD gap of the spatial-QP proxy suggests that future work could improve soft-indexing temperature schedules or add block-edge losses to tighten visual fidelity without losing differentiability.
If the transfer property holds for other non-differentiable modules (entropy coding, mode decision), the same proxy pattern could open gradient-based control of those modules as well.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid engineering paper that makes H.264 QP differentiable via soft-indexed DCVC-FM proxies and shows real-codec BD-rate gains; the transfer premise is load-bearing but not fatally unmeasured.

read the letter

The useful takeaway is that they turn a variable-rate learned codec into a QP-differentiable H.264 intra proxy (global and macroblock-level) via soft indexing, freeze it, and train AQ nets that deliver measurable rate-task gains when the resulting QPs are re-encoded with real JM 19.0. That is concrete and immediately usable for anyone still stuck with H.264 hardware.

What is new is the soft-index mapping from codec QP to the learned scaling factors, the explicit fidelity numbers (global-QP BD-rate gap only –4.75 % PSNR; spatial larger at –15 %), and the end-to-end AQ results on detection, segmentation and MS-SSIM. They do the right evaluation order: train against the proxy, report everything on the real codec. The RD curves, visual examples and BD-rate numbers (up to –17 % on Pascal VOC segmentation, –15 % on MS-SSIM) are clean enough to trust on the tested sets. Circularity is low; the proxy is matched to independent JM reconstructions.

Soft spots are real but proportionate. Soft-index temperature, the empirical α multiplier, the four anchor QPs and five uniform spatial distributions are hand-chosen and never ablated. Spatial-proxy fidelity is noticeably weaker, and the paper never quantifies how much of the AQ gain is residual proxy bias versus genuine H.264 improvement (no proxy-vs-real rate-task comparison on the same QP maps, no matched-bitrate heuristic baseline). That transfer gap is the load-bearing premise, yet the authors themselves note that perfect reconstruction is not required, and the real-codec curves still move in the right direction. Code is not released, so reproducibility is limited. Everything is still intra-only.

This is for people who need practical task-aware control inside the existing H.264 ecosystem, not for anyone looking for a new coding theory. The math is straightforward, the data support the claims that are made, and the citation pattern is honest. I would send it to peer review; a referee can demand the missing transfer ablation and code without killing the contribution. Worth a look if you care about codec proxies or rate-task optimization.

Referee Report

4 major / 0 minor

Summary. The paper proposes a differentiable proxy for the H.264 intra codec, built on a variable-rate learned compression model (DCVC-FM) and made QP-differentiable via soft indexing over learned scaling factors. Proxies are trained to match H.264 rate-distortion behavior under global-QP (one QP per image) and spatial-QP (macroblock-level) settings, with fidelity reported via RD curves and BD-rate on Kodak. The frozen proxies then supervise lightweight AQ networks that predict QP (or QP maps) for perceptual (MS-SSIM) and machine-vision objectives (detection, segmentation). Final evaluation re-encodes the predicted QPs with JM 19.0 and reports BD-rate savings of up to 17.12% (semantic segmentation) and 15.30% (MS-SSIM) versus fixed-QP H.264 baselines.

Significance. If the proxy-to-real transfer holds, the work offers a practical route to gradient-based adaptive quantization for a still-ubiquitous standard codec without modifying the codec itself. Explicit characterization of proxy fidelity (global and spatial) is a genuine contribution relative to prior proxy literature that often treats the proxy as an unexamined black box. Evaluation on the real JM encoder for the reported BD-rates, coverage of both perceptual and task metrics, and the spatial-QP formulation are strengths. The approach is reproducible in principle from the described architecture and training protocol, though code is not supplied in the manuscript.

major comments (4)

Section II-B and Eq. (9): the central claim rests on gradients from the frozen proxy producing QP maps that improve the real non-differentiable H.264 encoder. All AQ networks are optimized against proxy rate and reconstruction, then evaluated only after re-encoding with JM 19.0. There is no ablation that (a) scores the identical QP maps under the proxy’s own rate-task metric versus the real codec, or (b) compares proxy-optimized maps against non-proxy baselines (heuristic/random spatial QP of matched bitrate). Without this transfer measurement, the reported BD-rate gains (up to −17.12% / −15.30%) cannot be cleanly attributed to genuine H.264 improvement rather than residual proxy bias.
Section III-A.2 and Fig. 4: the spatial-QP proxy itself shows a large RD mismatch to H.264 (BD-rate −15.01% PSNR, −12.91% MS-SSIM), substantially worse than the global-QP proxy (−4.75% / −2.56%). The conclusion asserts that “highly accurate proxy reconstruction is not strictly necessary,” yet the largest AQ gains are claimed precisely under the spatial setting. The manuscript needs either a quantitative bound on how much mismatch still yields useful gradients or an experiment showing that AQ performance degrades gracefully as proxy fidelity is deliberately reduced.
Section III-B: all rate-task curves use only fixed-QP JM 19.0 as baseline. Standard H.264 already supports macroblock-level QP adaptation and rate-control heuristics; without comparison to those (or to prior proxy/AQ methods cited in the introduction), it is unclear how much of the reported gain is due to the learned proxy framework versus simply allowing any form of content-adaptive QP.
Section II-A: free parameters that directly shape the proxy (soft-index temperature τ, empirical multiplier α on λ, the four global anchor QPs, the five uniform spatial distributions, and the Kodak-based BPP–PSNR nearest-neighbor mapping) are fixed without sensitivity analysis. Because the soft-assignment (Eqs. 4–5) is the sole mechanism that makes the proxy differentiable w.r.t. QP, at least an ablation on τ and on the anchor set is needed to establish that the reported fidelity and AQ gains are not brittle to these choices.

Circularity Check

0 steps flagged

No circularity: proxy is supervised against independent H.264 (JM 19.0) outputs; AQ networks are optimized on the frozen proxy but evaluated exclusively on the real non-differentiable codec.

full rationale

The derivation chain is self-contained and non-circular. The proxy P_ heta is trained by minimizing a supervised loss (Eq. 6) that matches its reconstruction and rate to those of the external H.264 reference software C (JM 19.0) under sampled global-QP and spatial-QP configurations; the soft-indexing mechanism (Eqs. 4–5) merely relaxes discrete index selection for differentiability and does not define the target RD behavior. The subsequent AQ stage freezes the proxy and optimizes heta-independent networks heta against the proxy’s rate-task objective (Eq. 9), then re-encodes the resulting QP maps with the real codec for all reported BD-rate numbers. No equation equates a claimed prediction to a fitted input by construction, no uniqueness theorem is imported from the authors, and the sole base model (DCVC-FM) is an external citation. Empirical gaps between proxy and H.264 (BD-rate –4.75 % / –15.01 % PSNR) and the transfer of gradients are measured rather than assumed definitional; the paper’s own statement that “highly accurate proxy reconstruction is not strictly necessary” is an empirical observation, not a tautology. Consequently the reported rate-task gains cannot be reduced to the training inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 2 invented entities

The central empirical claims rest on standard codec and learning assumptions plus a handful of hand-chosen anchors, multipliers, and network designs that are not derived from first principles. No new physical entities are postulated; the free parameters and domain assumptions listed below are the main degrees of freedom that must hold for the reported transfer from proxy to real H.264 to succeed.

free parameters (5)

soft-index temperature τ
Controls sharpness of the continuous assignment over discrete scaling factors; chosen by hand and never ablated.
empirical multiplier α on Lagrange multiplier λ
Introduced ad hoc 'to better match some QP operating points' (Eq. 7); fitted rather than derived.
anchor QP levels {35,40,45,51} and five uniform spatial distributions
Hand-selected discrete operating points that define the entire training distribution of the proxy.
rate-task trade-off β and its modulation parameters w,b,α=0.5
Selected per task/model size to obtain comparable bitrates; free hyperparameters of the AQ objective.
proxy-to-QP mapping via nearest BPP-PSNR neighbor on Kodak
Empirical discrete mapping from codec QP to proxy control index; dataset-dependent and fixed after measurement.

axioms (3)

domain assumption A neural compressor trained to match H.264 reconstructions and bitrates under a finite set of QP anchors will supply useful gradients for QP optimization on unseen images and QPs.
Core transfer assumption stated in Sections II-A and II-B; never proved, only empirically checked on Kodak/COCO/VOC.
ad hoc to paper Soft assignment over learned scaling factors (Eqs. 4-5) is a sufficiently accurate differentiable surrogate for discrete QP selection inside H.264.
Introduced specifically for this proxy; inspired by soft-to-hard VQ but not shown to be unique or optimal.
domain assumption JM 19.0 H.264 reference software in YUV420/444 is an adequate stand-in for real-world H.264 deployments.
Standard practice in the field; used for all ground-truth targets.

invented entities (2)

soft-indexed H.264 proxy (global-QP and spatial-QP variants) no independent evidence
purpose: Provide a differentiable surrogate of H.264 rate and reconstruction so that AQ networks can be trained end-to-end.
New composite model built on DCVC-FM; independent evidence is limited to the reported RD matching curves on Kodak.
proxy-based AQ networks (global CNN and spatial FCN) no independent evidence
purpose: Predict image-level or macroblock-level QP maps optimized for a downstream task loss.
Standard lightweight architectures specialized for this control problem; no external validation beyond the paper's own rate-task plots.

pith-pipeline@v1.1.0-grok45 · 13663 in / 2801 out tokens · 39926 ms · 2026-07-14T11:23:14.459635+00:00 · methodology

0 comments

read the original abstract

H.264 has been the most widely used video coding format for the past two decades due to its relative simplicity, efficiency, and wide availability of software and hardware implementations. However, optimizing codec parameters such as the quantization parameter (QP) for specific objectives (e.g., perceptual quality or machine vision tasks) is challenging due to the non-differentiable nature of standard video codecs. While differentiable proxies have recently been used to enable gradient-based optimization around standard codecs, their fidelity to the target codec is rarely explicitly characterized. In this paper, we propose a differentiable proxy learning method for H.264 intra codec to enable adaptive quantization control. Built upon a variable-rate learned compression model, the proposed proxy is made differentiable with respect to codec QP through a soft-indexing mechanism. It is then trained to approximate the rate-distortion behavior of H.264 under two quantization settings: global-QP, which uses one QP per image, and spatial-QP, which assigns QPs at the macroblock level. Using the frozen trained proxy, we develop a proxy-based adaptive quantization (AQ) framework for both perceptual optimization and machine vision tasks. Experimental results demonstrate that the proposed proxies closely approximate the rate-distortion behavior of H.264 intra codec. The resulting proxy-based AQ framework consistently improves rate-task trade-offs over fixed-QP H.264 baselines, achieving BD-rate reduction of up to 17.12% for semantic segmentation and 15.30% for MS-SSIM.

Figures

Figures reproduced from arXiv: 2607.10478 by Ivan V. Baji\'c, Qihan Xu.

**Figure 2.** Figure 2: Proxy-based adaptive quantization control framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Structure of the AQ network for adaptive spatial-QP [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 5.** Figure 5: The proxy reproduces artifact patterns that roughly fol [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

**Figure 5.** Figure 5: Visual comparison between H.264 and the proposed proxy under global-QP and spatial-QP settings. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 7.** Figure 7: Rate–MS-SSIM comparison on Kodak. 2) Results on adaptive spatial-QP control: For adaptive spatial-QP control, we train AQ networks on Pascal VOC for semantic segmentation using DeepLabv3 [19] with a ResNet50 backbone, and on Vimeo-90k for perceptual optimization using MS-SSIM as the objective. The corresponding rate– task results are summarized in [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗

**Figure 8.** Figure 8: Visual comparison of segmentation results. (a) Original image. (b) Ground truth. (c) Reconstructed image using adaptive [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗

**Figure 9.** Figure 9: Rate-Accuracy comparison on Pascal VOC using [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 2 linked inside Pith

[1]

Overview of the h. 264/avc video coding standard,

Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra, “Overview of the h. 264/avc video coding standard,”IEEE Transactions on circuits and systems for video technology, vol. 13, no. 7, pp. 560–576, 2003

2003
[2]

Differentiable bit-rate estimation for neural-based video codec enhancement,

Amir Said, Manish Kumar Singh, and Reza Pourreza, “Differentiable bit-rate estimation for neural-based video codec enhancement,” in2022 Picture Coding Symposium (PCS), San Jose, CA, USA, Dec. 2022, pp. 379–383, IEEE

2022
[3]

Sandwiched compression: Repurposing standard codecs with neural network wrappers,

Onur G. Guleryuz, Philip A. Chou, Berivan Isik, Hugues Hoppe, Danhang Tang, Ruofei Du, Jonathan Taylor, Philip Davidson, and Sean Fanello, “Sandwiched compression: Repurposing standard codecs with neural network wrappers,”arXiv preprint arXiv:2402.05887, 2024

Pith/arXiv arXiv 2024
[4]

Deep network-based adaptive quantization for practical video coding,

Shuai Huo, Hewei Liu, Jiawen Gu, Dengchao Jin, Meng Lei, Bo Huang, and Chao Zhou, “Deep network-based adaptive quantization for practical video coding,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

2025
[5]

En- hanced Standard Compatible Image Compression Framework Based on Auxiliary Codec Networks,

Hanbin Son, Taeoh Kim, Hyeongmin Lee, and Sangyoun Lee, “En- hanced Standard Compatible Image Compression Framework Based on Auxiliary Codec Networks,”IEEE Transactions on Image Processing, vol. 31, pp. 664–677, 2022

2022
[6]

Codec-Simulation Network for Joint Optimization of Video Coding With Pre- and Post-Processing,

Kaitian Qiu, Lu Yu, and Daowen Li, “Codec-Simulation Network for Joint Optimization of Video Coding With Pre- and Post-Processing,” IEEE Open Journal of Circuits and Systems, vol. 2, pp. 648–659, 2021

2021
[7]

Deep video codec control for vision models,

Christoph Reich, Biplob Debnath, Deep Patel, Tim Prangemeier, Daniel Cremers, and Srimat Chakradhar, “Deep video codec control for vision models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) workshops, June 2024, pp. 5732–5741

2024
[8]

Joint au- toregressive and hierarchical priors for learned image compression,

David Minnen, Johannes Ball ´e, and George D Toderici, “Joint au- toregressive and hierarchical priors for learned image compression,” Advances in neural information processing systems, vol. 31, 2018

2018
[9]

Variational image compression with a scale hyperprior,

Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston, “Variational image compression with a scale hyperprior,” inInternational Conference on Learning Representations, 2018

2018
[10]

Perceptual Video Compression with Neural Wrapping,

Muhammad Umar Karim Khan, Aaron Chadha, Mohammad Ashraful Anam, and Yiannis Andreopoulos, “Perceptual Video Compression with Neural Wrapping,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17743–17754

2025
[11]

Preprocessing Enhanced Image Compression for Machine Vision,

Guo Lu, Xingtong Ge, Tianxiong Zhong, Qiang Hu, and Jing Geng, “Preprocessing Enhanced Image Compression for Machine Vision,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 12, pp. 13556–13568, Dec. 2024

2024
[12]

How to exploit the transferability of learned image compression to conventional codecs,

Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien, “How to exploit the transferability of learned image compression to conventional codecs,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 16165–16174

2021
[13]

Neural video compression with feature modulation,

Jiahao Li, Bin Li, and Yan Lu, “Neural video compression with feature modulation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26099–26108

2024
[14]

Kodak lossless true color image suite,

“Kodak lossless true color image suite,” http://r0k.us/graphics/kodak/, Accessed: 2025-05-01

2025
[15]

Soft-to-hard vec- tor quantization for end-to-end learning compressible representations,

Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cav- igelli, Radu Timofte, Luca Benini, and Luc V Gool, “Soft-to-hard vec- tor quantization for end-to-end learning compressible representations,” Advances in neural information processing systems, vol. 30, 2017

2017
[16]

Video enhancement with task-oriented flow,

Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman, “Video enhancement with task-oriented flow,” inInternational Journal of Computer Vision (IJCV). 2019, vol. 127, pp. 1106–1125, Springer

2019
[17]

JM reference software version 19.0,

Joint Video Team (JVT), “JM reference software version 19.0,” https: //iphome.hhi.de/suehring/tml/download/, 2010, H.264/A VC Reference Software

2010
[18]

Yolov5 by ultralytics,

Glenn Jocher, Ayush Chaurasia, Jing Qiu, and Alex Stoken, “Yolov5 by ultralytics,” https://github.com/ultralytics/yolov5, 2022, Version 7.0, accessed 2026-04-13

2022
[19]

Rethinking atrous convolution for semantic image segmenta- tion,

Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam, “Rethinking atrous convolution for semantic image segmenta- tion,”arXiv preprint arXiv:1706.05587, 2017

Pith/arXiv arXiv 2017

[1] [1]

Overview of the h. 264/avc video coding standard,

Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra, “Overview of the h. 264/avc video coding standard,”IEEE Transactions on circuits and systems for video technology, vol. 13, no. 7, pp. 560–576, 2003

2003

[2] [2]

Differentiable bit-rate estimation for neural-based video codec enhancement,

Amir Said, Manish Kumar Singh, and Reza Pourreza, “Differentiable bit-rate estimation for neural-based video codec enhancement,” in2022 Picture Coding Symposium (PCS), San Jose, CA, USA, Dec. 2022, pp. 379–383, IEEE

2022

[3] [3]

Sandwiched compression: Repurposing standard codecs with neural network wrappers,

Onur G. Guleryuz, Philip A. Chou, Berivan Isik, Hugues Hoppe, Danhang Tang, Ruofei Du, Jonathan Taylor, Philip Davidson, and Sean Fanello, “Sandwiched compression: Repurposing standard codecs with neural network wrappers,”arXiv preprint arXiv:2402.05887, 2024

Pith/arXiv arXiv 2024

[4] [4]

Deep network-based adaptive quantization for practical video coding,

Shuai Huo, Hewei Liu, Jiawen Gu, Dengchao Jin, Meng Lei, Bo Huang, and Chao Zhou, “Deep network-based adaptive quantization for practical video coding,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

2025

[5] [5]

En- hanced Standard Compatible Image Compression Framework Based on Auxiliary Codec Networks,

Hanbin Son, Taeoh Kim, Hyeongmin Lee, and Sangyoun Lee, “En- hanced Standard Compatible Image Compression Framework Based on Auxiliary Codec Networks,”IEEE Transactions on Image Processing, vol. 31, pp. 664–677, 2022

2022

[6] [6]

Codec-Simulation Network for Joint Optimization of Video Coding With Pre- and Post-Processing,

Kaitian Qiu, Lu Yu, and Daowen Li, “Codec-Simulation Network for Joint Optimization of Video Coding With Pre- and Post-Processing,” IEEE Open Journal of Circuits and Systems, vol. 2, pp. 648–659, 2021

2021

[7] [7]

Deep video codec control for vision models,

Christoph Reich, Biplob Debnath, Deep Patel, Tim Prangemeier, Daniel Cremers, and Srimat Chakradhar, “Deep video codec control for vision models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) workshops, June 2024, pp. 5732–5741

2024

[8] [8]

Joint au- toregressive and hierarchical priors for learned image compression,

David Minnen, Johannes Ball ´e, and George D Toderici, “Joint au- toregressive and hierarchical priors for learned image compression,” Advances in neural information processing systems, vol. 31, 2018

2018

[9] [9]

Variational image compression with a scale hyperprior,

Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston, “Variational image compression with a scale hyperprior,” inInternational Conference on Learning Representations, 2018

2018

[10] [10]

Perceptual Video Compression with Neural Wrapping,

Muhammad Umar Karim Khan, Aaron Chadha, Mohammad Ashraful Anam, and Yiannis Andreopoulos, “Perceptual Video Compression with Neural Wrapping,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17743–17754

2025

[11] [11]

Preprocessing Enhanced Image Compression for Machine Vision,

Guo Lu, Xingtong Ge, Tianxiong Zhong, Qiang Hu, and Jing Geng, “Preprocessing Enhanced Image Compression for Machine Vision,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 12, pp. 13556–13568, Dec. 2024

2024

[12] [12]

How to exploit the transferability of learned image compression to conventional codecs,

Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien, “How to exploit the transferability of learned image compression to conventional codecs,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 16165–16174

2021

[13] [13]

Neural video compression with feature modulation,

Jiahao Li, Bin Li, and Yan Lu, “Neural video compression with feature modulation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26099–26108

2024

[14] [14]

Kodak lossless true color image suite,

“Kodak lossless true color image suite,” http://r0k.us/graphics/kodak/, Accessed: 2025-05-01

2025

[15] [15]

Soft-to-hard vec- tor quantization for end-to-end learning compressible representations,

Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cav- igelli, Radu Timofte, Luca Benini, and Luc V Gool, “Soft-to-hard vec- tor quantization for end-to-end learning compressible representations,” Advances in neural information processing systems, vol. 30, 2017

2017

[16] [16]

Video enhancement with task-oriented flow,

Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman, “Video enhancement with task-oriented flow,” inInternational Journal of Computer Vision (IJCV). 2019, vol. 127, pp. 1106–1125, Springer

2019

[17] [17]

JM reference software version 19.0,

Joint Video Team (JVT), “JM reference software version 19.0,” https: //iphome.hhi.de/suehring/tml/download/, 2010, H.264/A VC Reference Software

2010

[18] [18]

Yolov5 by ultralytics,

Glenn Jocher, Ayush Chaurasia, Jing Qiu, and Alex Stoken, “Yolov5 by ultralytics,” https://github.com/ultralytics/yolov5, 2022, Version 7.0, accessed 2026-04-13

2022

[19] [19]

Rethinking atrous convolution for semantic image segmenta- tion,

Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam, “Rethinking atrous convolution for semantic image segmenta- tion,”arXiv preprint arXiv:1706.05587, 2017

Pith/arXiv arXiv 2017