pith. sign in

arxiv: 2606.21764 · v1 · pith:XQLOUUE7new · submitted 2026-06-19 · 💻 cs.CV

Motion-Aware Reinforcement Learning For Object Localization

Prithvi Raj Singh , Satyendra Singh This is my paper

Pith reviewed 2026-06-26 14:08 UTC · model grok-4.3

classification 💻 cs.CV
keywords reinforcement learningobject detectionbounding box refinementmotion priorPPOcomputer visionaction smoothness
0
0 comments X

The pith

A constant-velocity motion prior added to a PPO bounding-box refinement agent raises IoU success by up to 1.1 points on VOC.

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

The paper introduces MARLNet, a reinforcement learning agent that refines object bounding boxes by incorporating a constant-velocity kinematic prediction into its observation vector and an action smoothness penalty into its reward. This design is tested against plain PPO on Pascal VOC 2012 and VisDrone 2019, where the motion signal is shown to reduce overshooting and produce higher rates of successful detections at IoU of at least 0.5. The authors also isolate a reward interference that collapses the termination trigger when a velocity-deviation term is paired with an absolute IoU term, and they replace it with the smoothness penalty to restore stable training. They further document a representational ceiling that appears when the refinement agent shares its backbone features with the initial detector. These elements together demonstrate how motion information can stabilize and modestly improve iterative localization in still images.

Core claim

MARLNet is a PPO-based bounding-box refinement agent that incorporates a constant-velocity motion prior into the 268-dimensional observation state and an action smoothness penalty into the reward function. The agent learns a five-dimensional policy controlling coordinate adjustments and a binary termination trigger. On Pascal VOC 2012 it reaches detection success gains of up to +0.011 at IoU ≥ 0.5 when the physical regularization weight is 0.10; on VisDrone 2019 the gain is +0.007 at weight 0.70. The motion prior is credited with preventing the overshooting that causes plain PPO to regress on the success metric, while reward ablations confirm that the smoothness term avoids trigger collapse.

What carries the argument

The 268-dimensional observation that concatenates the current proposal, a constant-velocity kinematic prediction computed from the previous action, the previous action itself, and a 256-dimensional EfficientNet-B0 crop feature, together with the action smoothness penalty in the reward.

If this is right

  • The motion prior prevents overshooting during refinement and thereby raises IoU ≥ 0.5 success rates relative to unconstrained PPO on VOC.
  • On VisDrone the weaker base detector allows larger absolute gains for plain PPO, yet MARLNet still records a positive though smaller improvement.
  • Pairing a constant-velocity deviation penalty with an absolute IoU reward term produces trigger collapse; the action smoothness penalty eliminates this interference.
  • Crop-feature refinement agents that share a backbone with their base detector encounter a representational ceiling, as shown by the global-plus-local observation ablation.

Where Pith is reading between the lines

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

  • The same observation construction could be applied to video sequences where frame-to-frame motion is stronger and the constant-velocity assumption is more often valid.
  • Replacing the kinematic predictor with a learned motion model that handles acceleration would test whether the reported gains are tied specifically to the constant-velocity choice.
  • The reward-interference finding may generalize to other RL localization tasks that combine geometric and appearance-based reward terms.

Load-bearing premise

The constant-velocity kinematic prediction computed from the previous action remains a useful signal even when the base detector's initial proposals are noisy or when object motion deviates from constant velocity.

What would settle it

Running the identical MARLNet agent on a test set of objects whose motion is strongly non-constant (for example, rapidly accelerating vehicles filmed at high frame rate) and measuring whether the reported gains over plain PPO disappear.

Figures

Figures reproduced from arXiv: 2606.21764 by Prithvi Raj Singh, Satyendra Singh.

Figure 1
Figure 1. Figure 1: Overview of the proposed motion-aware reinforcement learning framework for bounding box refinement. Given an initial object hypothesis from a detector, visual features are extracted from the cropped region and passed to an actor–critic agent. The agent predicts a bounding box adjustment and a stop signal. The environment evaluates the resulting IoU improvement and applies a motion-based consistency penalty… view at source ↗
Figure 2
Figure 2. Figure 2: Inference pipeline of MARLNet. Given an input image, a fine-tuned Faster R-CNN provides an initial bounding-box hypothesis. At each refinement step, the current box defines a crop that is processed by a frozen EfficientNet-B0 backbone to produce a 256-dimensional feature vector ft. This is concatenated with the current box coordinates, a constant-velocity motion prediction, and the previous action to form … view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of refinement process by MARLNet on random images from the VOC validation split. Each row shows the initial detector box (red) and the final refined box (blue) for a single image. The agent learns to make small, precise adjustments that often improve IoU with the ground-truth annotation (green). The smoothness penalty encourages consistent refinement trajectories, as evidenced by the relativel… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of refinement process by MARLNet on random images from the VisDrone 2019 validation split. Each group of three columns shows the initial detector box (red), the final refined box (blue), and the ground-truth box (green). MARLNet successfully refines some marginally-aligned boxes over the 0.5 IoU threshold, but does not systematically correct localization errors, consistent with the feature-inf… view at source ↗
Figure 5
Figure 5. Figure 5: Training dynamics comparing the two motion prior formulations over 70 epochs. (a) Mean episodic reward: the smoothness prior (blue, solid) grows steadily, while the CV-deviation prior (red, dashed) remains flat near zero throughout training. (b) Mean training IoU: the smoothness prior rises to 0.51 before the run was stopped at epoch 69, consistently approaching the detector initialization IoU of 0.487 (do… view at source ↗
Figure 6
Figure 6. Figure 6: Inference latency of MARLNet on GPU and CPU across both datasets. The policy forward pass is negligible compared to feature extraction and the detector, confirming that the RL component does not introduce a significant computational bottleneck. 5 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
read the original abstract

We present MARLNet (Motion-Aware Reinforcement Learning Network), a PPO-based bounding-box refinement agent that incorporates a constant-velocity motion prior into the observation state and an action smoothness penalty into the reward function. The agent operates on 268-dimensional observations encoding the current proposal, a kinematic prediction, the previous action, and a 256-dimensional EfficientNet-B0 crop feature, and learns a five-dimensional policy controlling coordinate adjustments and a binary termination trigger. Evaluated on Pascal VOC 2012 and VisDrone 2019, MARLNet trains stably across all regularization strengths tested and achieves consistent gains in detection success rate at $\text{IoU} \geq 0.5$: up to $+0.011$ on VOC ($\lambda_\text{phys}{=}0.10$), where the motion prior prevents the overshooting that causes plain PPO to regress on this metric, and $+0.007$ on VisDrone ($\lambda_\text{phys}{=}0.70$), where unconstrained PPO achieves a larger gain ($+0.025$) owing to the weaker base detector. Through reward design ablations and training dynamics analysis, we identify a reward interference in which combining a constant-velocity deviation penalty with an absolute IoU term causes trigger collapse, and show that replacing it with the action smoothness penalty resolves this failure. We further characterize a representational ceiling facing crop-feature refinement agents that share a backbone with their base detector, confirmed through a global-plus-local observation ablation. Project page: https://prithviraj97.github.io/marl-net

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents MARLNet, a PPO-based bounding-box refinement agent for object detection that augments the 268-dimensional observation state with a constant-velocity kinematic prediction derived from the prior action and incorporates an action smoothness penalty in the reward. It reports empirical gains in IoU≥0.5 success rate of up to +0.011 on Pascal VOC 2012 (at λ_phys=0.10) and +0.007 on VisDrone 2019 (at λ_phys=0.70), attributes the VOC improvement to the motion prior preventing overshooting seen in plain PPO, identifies a reward-interference failure mode (trigger collapse) via ablations on constant-velocity deviation penalties versus smoothness penalties, and notes a representational ceiling for agents sharing a backbone with the base detector.

Significance. If the reported gains prove robust, the work offers a concrete demonstration that kinematic priors in the observation and targeted reward penalties can stabilize RL refinement agents and avoid specific failure modes such as overshooting or trigger collapse; the reward-design ablations and training-dynamics analysis constitute a useful contribution to RL-for-vision practice. The project page further supports reproducibility. The modest effect sizes and the fact that unconstrained PPO outperforms on VisDrone nevertheless limit the broader significance.

major comments (2)
  1. [Observation construction and evaluation on VOC/VisDrone] The central claim that the motion prior prevents overshooting (and thereby produces the +0.011 VOC gain) rests on the constant-velocity kinematic prediction remaining informative. No experiment tests this channel when initial proposals are noisy or when object trajectories deviate from constant velocity, which directly affects attribution of the reported numeric lifts.
  2. [Quantitative results and ablations] The headline numeric results (+0.011 on VOC, +0.007 on VisDrone) are presented as point estimates with no error bars, no standard deviations across random seeds, and no statistical significance tests. This is load-bearing for the claim of “consistent gains” and “stable training across all regularization strengths.”
minor comments (2)
  1. [Method] The 268-dimensional observation breakdown (current proposal, kinematic prediction, previous action, 256-dim EfficientNet crop) would benefit from an explicit per-component dimension table for immediate clarity.
  2. [Abstract] The abstract states that unconstrained PPO achieves a larger gain (+0.025) on VisDrone; a brief discussion of why the motion prior is still presented as beneficial in that regime would help readers interpret the cross-dataset pattern.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below, providing clarifications and committing to revisions where appropriate to strengthen the manuscript.

read point-by-point responses

  1. Referee: The central claim that the motion prior prevents overshooting (and thereby produces the +0.011 VOC gain) rests on the constant-velocity kinematic prediction remaining informative. No experiment tests this channel when initial proposals are noisy or when object trajectories deviate from constant velocity, which directly affects attribution of the reported numeric lifts.

    Authors: The reported results are obtained on Pascal VOC 2012 and VisDrone 2019 using proposals from a base detector, which inherently include noise, and real object trajectories that are not strictly constant-velocity. Attribution of the VOC gain to the motion prior is supported by the training dynamics analysis showing that plain PPO overshoots while the augmented agent does not. We nevertheless agree that the claim would be strengthened by explicit tests of the kinematic channel under controlled noise levels or non-constant-velocity conditions. In the revised manuscript we will add an ablation that perturbs initial proposal noise and trajectory assumptions to directly evaluate when the constant-velocity prediction remains informative. revision: yes

  2. Referee: The headline numeric results (+0.011 on VOC, +0.007 on VisDrone) are presented as point estimates with no error bars, no standard deviations across random seeds, and no statistical significance tests. This is load-bearing for the claim of “consistent gains” and “stable training across all regularization strengths.”

    Authors: We acknowledge that the headline numbers are reported as single-run point estimates. While the training-dynamics figures demonstrate stability across regularization strengths, the absence of variability measures and significance tests weakens the support for claims of consistent gains. In the revision we will rerun the key experiments with multiple random seeds, report means and standard deviations, and include statistical significance tests for the reported improvements. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical RL results on fixed datasets

full rationale

The paper describes a PPO-based agent whose 268-dim observation includes a constant-velocity kinematic extrapolation and whose reward includes an action-smoothness term. All reported metrics (+0.011 IoU gain on VOC, +0.007 on VisDrone) are measured outcomes of training runs on standard dataset splits; none are quantities defined in terms of λ_phys or obtained by algebraic rearrangement of the observation vector. No self-citation chain, uniqueness theorem, or fitted-parameter-renamed-as-prediction appears in the derivation. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central empirical claim rests on standard PPO convergence assumptions, the validity of the constant-velocity motion model for the chosen datasets, and the independence of the refinement stage from the base detector's training. No new physical constants or particles are introduced.

free parameters (1)
  • λ_phys = 0.10 / 0.70
    Regularization coefficient weighting the physical (motion) term in the reward; tested at 0.10 on VOC and 0.70 on VisDrone and described as chosen after sweeps.
axioms (2)
  • domain assumption PPO policy-gradient updates converge to a useful policy when the observation includes a constant-velocity forecast derived from the previous action.
    Invoked when the 268-dimensional state vector is defined and when the motion prior is credited with preventing overshoot.
  • domain assumption The EfficientNet-B0 crop feature extracted from the current proposal remains informative even when the backbone is shared with the base detector.
    Underlying the global-plus-local observation ablation that identifies the representational ceiling.

pith-pipeline@v0.9.1-grok · 5807 in / 1669 out tokens · 20495 ms · 2026-06-26T14:08:28.167670+00:00 · methodology

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]

    Beyond kalman filters: deep learning-based filters for improved object tracking

    Momir Adžemović, Predrag Tadić, Andrija Petrović, and Mladen Nikolić. Beyond kalman filters: deep learning-based filters for improved object tracking. Machine Vision and Applications, 36(1):20, 2025. 3

    2025

  2. [2]

    Simple online and re- altime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and re- altime tracking. In2016 IEEE international confer- ence on image processing (ICIP), pages 3464–3468. Ieee, 2016. 1, 3

    2016

  3. [3]

    Kinematic 3d object detection in monocular video

    Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3d object detection in monocular video. InEuropean Conference on Computer Vision, pages 135–152. Springer, 2020. 3

    2020

  4. [4]

    Active object localization with deep reinforcement learning

    Juan C Caicedo and Svetlana Lazebnik. Active object localization with deep reinforcement learning. InProceedings of the IEEE international conference on computer vision, pages 2488–2496, 2015. 1, 3

    2015

  5. [5]

    Visible and clear: Finding tiny objects in difference map

    Bing Cao, Haiyu Yao, Pengfei Zhu, and Qinghua Hu. Visible and clear: Finding tiny objects in difference map. InEuropean Conference on Computer Vision, pages 1–18. Springer, 2024. 8

    2024

  6. [6]

    Enhancing your trained detrs with box refinement.arXiv preprint arXiv:2307.11828, 2023

    Yiqun Chen, Qiang Chen, Peize Sun, Shoufa Chen, Jingdong Wang, and Jian Cheng. Enhancing your trained detrs with box refinement.arXiv preprint arXiv:2307.11828, 2023. 1, 2

    arXiv 2023

  7. [7]

    Iterative multiple bounding-box refine- ments for visual tracking.Journal of Imaging, 8(3): 61, 2022

    Giorgio Cruciata, Liliana Lo Presti, and Marco La Cascia. Iterative multiple bounding-box refine- ments for visual tracking.Journal of Imaging, 8(3): 61, 2022. 2

    2022

  8. [8]

    Visdrone-det2019: The vision meets drone object detection in image challenge results

    Dawei Du, Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Lin, Qinghua Hu, Tao Peng, Jiayu Zheng, Xinyao Wang, Yue Zhang, et al. Visdrone-det2019: The vision meets drone object detection in image challenge results. InProceedings of the IEEE/CVF 14 international conference on computer vision work- shops, pages 0–0, 2019. 8

    2019

  9. [9]

    Visual foresight: Model-based deep reinforcement learning for vision-based robotic control.arXiv preprint arXiv:1812.00568, 2018

    Frederik Ebert, Chelsea Finn, Sudeep Dasari, et al. Visual foresight: Model-based deep reinforcement learning for vision-based robotic control.arXiv preprint arXiv:1812.00568, 2018. 3

    Pith/arXiv arXiv 2018

  10. [10]

    The pascal visual object classes (voc) challenge.Inter- national journal of computer vision, 88(2):303–338,

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge.Inter- national journal of computer vision, 88(2):303–338,

  11. [11]

    Pm- track: multi-object tracking with motion-aware

    Xu Guo, Yujin Zheng, and Dingwen Wang. Pm- track: multi-object tracking with motion-aware. In Proceedings of the Asian Conference on Computer Vision, pages 3091–3106, 2024. 1, 3

    2024

  12. [12]

    Deep residual learning for image recog- nition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770– 778, 2016. 5

    2016

  13. [13]

    Bbre- finement: A universal scheme to improve the preci- sion of box object detectors.Applied Sciences, 12 (7):3402, 2022

    Petr Hurtik, Marek Vajgl, and David Hynar. Bbre- finement: A universal scheme to improve the preci- sion of box object detectors.Applied Sciences, 12 (7):3402, 2022. 1, 2

    2022

  14. [14]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014. 7, 1

    Pith/arXiv arXiv 2014

  15. [15]

    Multi-stage reinforcement learning for object detection

    Jonas König, Simon Malberg, Martin Martens, Se- bastian Niehaus, Artus Krohn-Grimberghe, and Arunselvan Ramaswamy. Multi-stage reinforcement learning for object detection. InScience and Infor- mation Conference, pages 178–191. Springer, 2019. 3

    2019

  16. [16]

    Gradient-based regulariza- tion for action smoothness in robotic control with reinforcement learning

    I Lee, Hoang-Giang Cao, Cong-Tinh Dao, Yu-Cheng Chen, and I-Chen Wu. Gradient-based regulariza- tion for action smoothness in robotic control with reinforcement learning. In2024 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS), pages 603–610. IEEE, 2024. 3

    2024

  17. [17]

    Microsoft coco: Com- mon objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Com- mon objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014. 8

    2014

  18. [18]

    Kurl: A knowledge-guided reinforce- ment learning model for active object tracking

    Xin Liu, Jie Tan, Xiaoguang Ren, Weiya Ren, and Huadong Dai. Kurl: A knowledge-guided reinforce- ment learning model for active object tracking. In Asian conference on machine learning, pages 818–

  19. [19]

    Liu et al

    X. Liu et al. Traffic video object detection using mo- tion prior.arXiv preprint arXiv:2311.10092, 2023. 3

    arXiv 2023

  20. [20]

    Reinforcement learning for visual object detection

    Stefan Mathe, Aleksis Pirinen, and Cristian Smin- chisescu. Reinforcement learning for visual object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2894–2902, 2016. 3

    2016

  21. [21]

    Regularizing action policies for smooth control with reinforcement learning

    Siddharth Mysore, Bassel Mabsout, Renato Man- cuso, and Kate Saenko. Regularizing action policies for smooth control with reinforcement learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1810–1816. IEEE,

    2021

  22. [22]

    Visual object tracking in drone im- ages with deep reinforcement learning

    Sedat Ozer et al. Visual object tracking in drone im- ages with deep reinforcement learning. In2020 25th International Conference on Pattern Recognition (ICPR), pages 10082–10089. IEEE, 2021. 3

    2021

  23. [23]

    Deep reinforcement learning of region proposal networks for object detection

    Aleksis Pirinen and Cristian Sminchisescu. Deep reinforcement learning of region proposal networks for object detection. Inproceedings of the IEEE con- ference on computer vision and pattern recognition, pages 6945–6954, 2018. 3

    2018

  24. [24]

    Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information processing systems, 28, 2015

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information processing systems, 28, 2015. 2, 5, 7

    2015

  25. [25]

    High- dimensional continuous control using general- ized advantage estimation.arXiv preprint arXiv:1506.02438, 2015

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High- dimensional continuous control using general- ized advantage estimation.arXiv preprint arXiv:1506.02438, 2015. 7

    Pith/arXiv arXiv 2015

  26. [26]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhari- wal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017. 1, 5, 7

    Pith/arXiv arXiv 2017

  27. [27]

    Efficientnet: Rethink- ing model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethink- ing model scaling for convolutional neural networks. InInternational conference on machine learning, pages 6105–6114. PMLR, 2019. 2, 4, 5

    2019

  28. [28]

    Motiontrack: Learning motion predictor for multi- ple object tracking.Neural Networks, 179:106539,

    Changcheng Xiao, Qiong Cao, Yujie Zhong, Long Lan, Xiang Zhang, Zhigang Luo, and Dacheng Tao. Motiontrack: Learning motion predictor for multi- ple object tracking.Neural Networks, 179:106539,

  29. [29]

    Alpha-refine: Boosting track- ing performance by precise bounding box estima- tion

    Bin Yan, Xinyu Zhang, Dong Wang, Huchuan Lu, and Xiaoyun Yang. Alpha-refine: Boosting track- ing performance by precise bounding box estima- tion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5289–5298, 2021. 1, 2

    2021

  30. [30]

    Lpae-yolov8: lightweight aerial small object detection via lse-head and adaptive attention.Scientific Reports, 15(1): 44780, 2025

    Yi Zhao and Guiping Chen. Lpae-yolov8: lightweight aerial small object detection via lse-head and adaptive attention.Scientific Reports, 15(1): 44780, 2025. 8

    2025

  31. [31]

    Rein- forcenet: A reinforcement learning embedded object detection framework with region selection network

    Man Zhou, Rujing Wang, Chengjun Xie, Liu Liu, Rui Li, Fangyuan Wang, and Dengshan Li. Rein- forcenet: A reinforcement learning embedded object detection framework with region selection network. Neurocomputing, 443:369–379, 2021. 1, 3

    2021

  32. [32]

    Detection and tracking meet drones challenge.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 44(11):7380–7399, 2021

    Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 44(11):7380–7399, 2021. 2, 7 15 Motion-Aware Reinforcement Learning For Object Localization Supplementary Material

    2021

  33. [33]

    The shared MLP backbone comprises two fully con- nected layers of dimension 256 with ReLU activa- tions

    Implementation Details Network architecture and feature extraction. The shared MLP backbone comprises two fully con- nected layers of dimension 256 with ReLU activa- tions. The actor head outputs a 4-dimensional mean vectorµt, a globally learnable log-standard-deviation logσ∈R4, and a scalar stopping logit; the critic head outputs a scalar value estimate ...

  34. [34]

    More Results on VOC & VisDrone Tables 7 and 8 report per-category AP@0.5 for all methods on VOC and VisDrone respectively, com- plementing the aggregate metrics in the main paper

  35. [35]

    To address this limitation, we extend the frame- work by introducing alearnable motion modulethat predicts the next bounding box in a data-driven manner

    Mathematical Formulation of the Learnable Motion Model While the fixed constant–velocity assumption pro- vides a simple and effective motion prior, it cannot capture complex object dynamics or camera motion. To address this limitation, we extend the frame- work by introducing alearnable motion modulethat predicts the next bounding box in a data-driven man...

    arXiv 2011

  36. [36]

    The method requires no training and no ground-truth supervision at inference time

    Heuristic Refinement Policy We describe the greedy detector-rescore heuristic used as a non-learned baseline throughout the ex- periments. The method requires no training and no ground-truth supervision at inference time. It shares the same base detector as MARLNet, ensuring a fair comparison: both methods operate from identical initial proposals and use ...

  37. [37]

    Discussion on Results of Experi- ment 13.1. Reward Design Analysis The CV-penalty reward causes catastrophic failure.When the motion penalty is defined as deviation from the constant-velocity prediction – λphys∥bt+1−ˆbcv t+1∥2 – training collapses by epoch

  38. [38]

    Figure 5(b) shows mean training IoU degrading from 0.42 to 0.21, confirming complete localization failure

    The mean stopping probability¯pstop falls below 0.01 and never recovers, reducing the model to ran- dom per-step jitter over allTmax steps. Figure 5(b) shows mean training IoU degrading from 0.42 to 0.21, confirming complete localization failure. Algorithm 2Greedy Heuristic Refinement Input:Image I; initial detection(b0,s 0); detector D; max stepsT=8; sca...