Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Temporal motion evidence, not bigger networks, is what makes aerial person tracking work on milliwatt hardware.

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 · deepseek-v4-flash

2026-08-02 07:44 UTC pith:C6KPNIXA

load-bearing objection A genuinely clever systems paper with unusually honest ablations; the transfer claims rest on a synthetic-motion protocol that needs one more independent check. the 3 major comments →

arxiv 2607.16282 v1 pith:C6KPNIXA submitted 2026-07-09 cs.CV

Moving Like a Human: Ego-Motion-Normalized Temporal Signatures for Real-Time Aerial Person Tracking on Milliwatt-Class Hardware

Akbar Anbar Jafari , Cagri Ozcinar , Gholamreza Anbarjafari This is my paper
classification cs.CV
keywords ego-motion compensationtemporal signaturessmall-object detectionperson trackingint8 quantizationsynthetic training curriculumaerial videoedge deployment
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.

The paper tries to show that on milliwatt-class drone hardware, person tracking does not need a bigger network; it needs the right input. By removing the camera's own motion with classical geometry and handing a tiny detector three input channels — luminance plus two ego-motion-normalized residual-motion maps — a 22k-parameter stateless network can detect people barely visible as 10-60 pixel smudges. The evidence is temporal, but it is computed analytically, not learned. The system generalizes to real drone imagery where a single-frame detector fine-tuned on the same data collapses, while running at real-time speed on a small single-board computer.

Core claim

Temporal evidence for detecting small moving people belongs in the input representation, computed by estimating and removing ego-motion, not in learned temporal machinery or appearance capacity. The deployed detector is stateless: adding temporal-shift modules lowers accuracy once the input channels carry ego-motion-normalized motion, and removing the motion channels barely changes in-domain score but collapses held-out real drone imagery 6x. Trained on synthetic clips whose motion channels were generated by the deployed ego-motion code, the int8 detector reaches 0.694 AP25 in-domain and 0.444 on held-out real drone stills, versus 0.103 for a far larger single-frame detector fine-tuned on th

What carries the argument

The carrying mechanism is the ego-motion-normalized residual channel stack: each frame is warped by a 4-DoF similarity transform estimated from sparse optical-flow tracks, and the network receives luminance, residual-motion magnitude, and signed temporal difference. Under exact registration, static scenes vanish from the motion channels, so everything bright there is independently moving; registration noise becomes a structured nuisance the detector learns to discount. Because the same code generates these channels during training and deployment, the synthetic-to-real gap shrinks. Motion evidence is thereby supplied analytically for free, which is what lets a 7.6-MFLOP stateless detector rep

Load-bearing premise

The generalization result rests on the assumption that the hand-set synthetic-motion recipe — up to 4 px/frame drift, 0.008 rad/frame roll, 0.6% scale rate, and per-person lower-body shear applied to static imagery — reproduces the task-relevant motion statistics of real drone footage.

What would settle it

Measure the deployed detector on a sizable corpus of real drone video with independently human-labeled person boxes, especially at 10-60 px target heights, without converting frames through the synthetic pseudo-clip procedure; if its AP25 there does not clearly beat an identically evaluated single-frame detector, the transfer claim is refuted.

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

If this is right

  • Follow-me drones can track a person with a 22k-parameter detector plus classical geometry, at 31 FPS on milliwatt-class hardware, holding lock through occlusion.
  • Accuracy that transfers to real imagery comes from motion evidence, not appearance capacity: variants without motion channels collapse 6x on held-out real data, and fine-tuning a large single-frame detector on the same data does not save it.
  • Learned temporal modules are unnecessary and harmful once analytic motion channels are present; stateless inference is the right architecture.
  • int8 deployment is only free if calibrated correctly: moving-average range estimation silently halves accuracy, regression heads need true min-max ranges, and stateful graphs need propagated cache states.
  • A 1-D verifier of human-motion signatures can halve false positives on person-free footage, enabling a low false-alarm system.

Where Pith is reading between the lines

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

  • If the transfer claim holds, the same recipe could apply to other small moving objects on moving platforms — animals or vehicles — provided their motion signatures can be synthesized in training.
  • The dependency on a synthetic-motion curriculum means the headline numbers are only as trustworthy as the hand-set trajectory parameters; an independently labeled real benchmark with distant-regime ground truth would be the decisive test.
  • The paper acknowledges that a stationary person starves the motion channels; this implies a hybrid mode that re-activates appearance evidence when motion is absent could extend coverage beyond the current design.
  • The verifier's 'moves like a human' test is trained on walk cycles; testing whether gait-phase articulation is enough to reject other periodic movers, such as animals, would sharpen the claim about what the motion signature actually captures.

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

3 major / 4 minor

Summary. The paper proposes EMTS-Det, a five-stage aerial person-tracking pipeline for milliwatt-class hardware. Stage A estimates ego-motion as a 4-DoF similarity transform; Stage B forms ego-motion-normalized residual-motion channels; Stage C is a 21,941-parameter, 7.6-MFLOP stateless CenterNet-style detector; Stage D is a Kalman tracker in stabilized coordinates; Stage E is a 1-D CNN verifier of human motion. Training uses a synthetic-motion curriculum with motion channels generated by the deployed Stage A/B code. The central claim is that temporal evidence belongs in the input representation, computed analytically, rather than in learned temporal machinery, and that this representation transfers to real drone imagery. Evidence includes multi-seed ablations on a composite bench and a held-out VisDrone-DET split, int8 quantization-failure analysis, an on-device Raspberry Pi Zero 2W evaluation over 1,000 real UAV videos, and a field sequence.

Significance. If the central claim holds, the paper is significant: it demonstrates a 1,100x compute reduction over YOLOv8n while maintaining or improving small-person detection on real imagery, provides a concrete negative result for learned temporal modules (TSM) once motion channels are available, and documents reproducible int8 calibration failure modes for stateless and stateful detectors. The paper is disciplined in several ways: ablations are repeated over three seeds with standard deviations; YOLOv8n is fine-tuned on the identical training mixture; arithmetic is internally consistent; limitations are admitted; and the authors release scripts for training and on-device measurement. However, the load-bearing generalization evidence currently depends on synthetic motion applied to static VisDrone images, and the only genuine real-video accuracy table lacks a stated labeling protocol. These gaps must be closed or substantially reframed before the transfer claim can be accepted as demonstrated.

major comments (3)
  1. [Sec. 6.5, Sec. 4.0.2, Tables 4-5] The held-out VisDrone-DET generalization evidence is obtained by converting static images into 4-frame pseudo-clips using the authors' Stage A/B code with hand-set trajectory parameters: up to 4 px/frame drift, 0.008 rad/frame roll, 0.6% scale rate, and per-person lower-body shear. The motion channels therefore encode the authors' model of drone/person motion, not measured drone motion. The headline 0.444-vs-0.103 comparison and the 6x collapse of the luminance-only variant in Table 5 are conditional on this synthetic recipe reproducing the task-relevant statistics of real UAV footage. Limitation 2 explicitly concedes that a field-collected labeled benchmark is still needed. Please add a sensitivity analysis over the trajectory parameters, or better, evaluate on real videos with hand labels, before presenting 0.444 as evidence of transfer to real imagery.
  2. [Sec. 6.11, Table 10] The 1,000-video real-world evaluation on Raspberry Pi Zero 2W is the only accuracy result on genuine videos, but the labeling protocol is not stated. Who created the ground-truth boxes, how were tiny/distant persons annotated, were labels manual or machine-generated, and what quality control was used? Without this information the AP25/AP50/recall numbers in Table 10 cannot be audited; if labels are YOLO-derived, the comparison against YOLOv8n would be systematically biased. Please provide the full annotation protocol, including inter-annotator agreement if applicable, and ideally release the labels or a public benchmark.
  3. [Secs. 6.8, 6.12 and Limitation 4] The field-sequence lock recall of 97.9% is measured only on YOLO-confirmed frames, which the paper states are close/mid-range frames. The distant regime, which covers 78% of the sequence and is exactly the regime the method targets, has no independent ground truth; Limitation 4 says so explicitly. As a result, the claimed advantage of the temporal representation over single-frame detectors is not directly validated in the operating regime where the paper argues it matters most. Please either hand-label a sample of distant frames, or clearly separate the claim into 'validated on close/mid range' and 'unvalidated on distant range'.
minor comments (4)
  1. [Table 5] The YOLOv8n row is formatted as '0.886 0.8710.103', which is ambiguous; it should be '0.886 / 0.871 / 0.103'.
  2. [Fig. 4 / Table 4] The classical diff+blob baseline is reported as AP25=0.120 in Table 4 but the precision-recall legend in Fig. 4 shows AP=0.119. Please reconcile the thresholds or state that the difference is due to the operating point.
  3. [Abstract / body] Several places render 'UAV' as 'UA V' (e.g., abstract and Section 6.11). Please fix the spacing.
  4. [Sec. 6.8] The verified false-lock rate of ~720/hr is based on only 2 minutes of screen-admitted footage and is dominated by one dense-vegetation video. The paper notes this, but a small confidence interval or per-video breakdown would help calibrate the KPI.

Circularity Check

1 steps flagged

Partial circularity: the central transfer result on 'held-out real VisDrone' is measured on pseudo-video generated by the same synthetic-motion recipe used in training, so the temporal-evidence claim is partly self-referential.

specific steps
  1. fitted input called prediction [Sec. 4.0.2 (Tier-2 pseudo-video recipe) and Sec. 6.5 (held-out VisDrone-DET evaluation)]
    "Evaluating the deployed int8 model on this held-out split (each image converted to a 4-frame pseudo-clip through the deployed Stage A/B code; 9,484 person instances) yields AP 25=0.444 ... The result confirms the model generalizes to real drone imagery outside the composite generative process."

    The held-out split is not real temporal video: each still image is converted to a 4-frame pseudo-clip by the same Stage A/B code and the same synthetic similarity-trajectory recipe (drift up to 4 px/frame, roll 0.008 rad/frame, scale rate 0.6%, per-person lower-body shear) described in Sec. 4.0.2 for Tier-2 pretraining, and Tier-2 pseudo-video is present in every Phase-2 training batch. Thus the test-time temporal signature is generated by the same function that produced the training motion channels. The 0.444-vs-0.051 collapse for luminance-only variants demonstrates sensitivity to this recipe, not transfer to independently captured drone motion; the temporal part of the claimed generalization is self-referential.

full rationale

No self-citation chains or uniqueness imports appear in the reference list; the detector/verifier ablations are retrained from scratch with fixed seeds and compared symmetrically, which is not circular. The main circularity concern is evaluation construction: the headline transfer evidence on held-out VisDrone-DET is obtained by converting static images into pseudo-clips with the deployed Stage A/B code and the same hand-set synthetic-motion recipe used in Tier-2 training, so the temporal signal in the test set is produced by the same generative process the model was trained on. The paper itself concedes the close cousin of this issue in Limitation 2 ('The composite detection bench shares a generative process with training') and Limitation 4 (no hand-labeled ground truth for the distant regime). The independent real-corpus result (Table 10, 1,000 UAV videos, 0.462 AP25) partially mitigates the circularity, but its labeling protocol is unstated, so it does not fully audit the transfer claim. On balance the central derivation—ego-motion normalization, residual channels, stateless detector—is not circular in an equation-level sense; the partial circularity lies in the construction of the key held-out evaluation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on four hand-set constant groups (pseudo-motion test parameters, tracker gates/noise, operating thresholds, Pi scaling factors) and five domain assumptions. The most consequential free parameters are the pseudo-video trajectory settings that define the VisDrone-DET generalization protocol, since the primary transfer argument is measured under them. No new physical entities are postulated, and there are no self-citations in the circularity chain; the circular burden is structural (self-generated test regimes), not citation-based.

free parameters (4)
  • Pseudo-video trajectory parameters (VisDrone/Tier-2 evaluation) = up to 4 px/frame drift, 0.008 rad/frame roll, 0.6% scale rate, oscillating lower-body shear
    Hand-set constants define the synthetic motion regime under which the headline generalization split (VisDrone-DET, 0.444 AP25) and the luminance-collapse ablation (0.415 -> 0.051) are measured.
  • Auto-lock and verifier operating thresholds = tau_s=0.30, verifier=0.35, Stage-E every 8th frame, verifier threshold 0.5
    Chosen operating points determine the 720/hr false-lock KPI, the FP/frame table, and the lock-recall figures.
  • Tracker noise and association gates = Q=diag(1,1,4,4,0.25,0.25), R=diag(4,4,9,9), tau_0=0.5, gamma^2<9.21, re-lock s>0.65, coast gate gamma^2<16, 3-frame pers
    Hand-chosen constants shape reacquisition success (93.6% over 783 trials) and the scale-aware gate (Eq. 10).
  • Pi-class platform scaling factors = 12x for int8 conv, 3x for memory-bound OpenCV/NumPy stages; pessimistic 15x/4x
    Used to produce the pre-measurement budget of 29.6 ms/frame; later validated within 6% by the on-device measurement of 31.4 ms/frame.
axioms (5)
  • domain assumption A 4-DoF similarity transform estimated by LK+RANSAC (with phase-correlation fallback) registers the moving camera accurately enough that static-scene residual vanishes except on high-gradient edges, and residual parallax is small relative to the 8-px output stride.
    Sec. 3.2-3.3; the entire motion-channel representation depends on this registration accuracy.
  • domain assumption A 10-60 pixel person is beyond the reach of any single-frame appearance detector at useful accuracy in this budget.
    Sec. 1, asserted as a contention ('we contend none can'); motivates the architecture but is not demonstrated by the baseline set, which omits modern sub-100-MFLOP single-frame detectors.
  • domain assumption Motion channels computed by the deployed code on synthetic composites are distributionally close to real-footage motion channels (small sim-to-real gap).
    Sec. 2.5 and 4; the core transfer argument rests on this, and the VisDrone generalization split is itself generated under this assumption.
  • domain assumption YOLOv8n at conf>=0.55 provides valid pseudo-ground-truth for lock recall on close/mid-range field frames, and the distant regime can be left unlabeled.
    Sec. 6.8; the 97.9% lock recall is measured only on YOLO-confirmed frames; the authors admit the distant regime (78% of the sequence) has no independent ground truth (Limitation 4).
  • standard math Walking cadence 1.5-2.5 Hz is inside the Nyquist rate of 30 Hz sampling, and a 16-frame window spans roughly one gait cycle.
    Sec. 3.6; standard signal-processing arithmetic, reasonable.

pith-pipeline@v1.3.0-alltime-deepseek · 22124 in / 18303 out tokens · 179346 ms · 2026-08-02T07:44:14.560045+00:00 · methodology

0 comments
read the original abstract

Follow-me person tracking must run on the drone itself, where affordable companion computers offer only a few effective int8 GFLOP/s. At typical follow distances a person spans 10-60 pixels, indistinguishable from clutter and beyond the reach of single-frame appearance detectors. The missing evidence is temporal and belongs in the input representation, computed analytically, rather than in learned temporal machinery. EMTS-Det is a five-stage system that estimates ego-motion, converts each frame into ego-motion-normalized residual-motion channels, detects person centers with a 22k-parameter, 7.6-MFLOP network, tracks a locked target with a Kalman filter in stabilized coordinates, and verifies tracks with a 1-D convolutional classifier of human motion (ROC AUC 0.941). Training uses a synthetic-motion curriculum with motion channels generated by the deployed ego-motion code. Multi-seed ablations locate the value in generalization: on held-out VisDrone-DET a luminance-only variant collapses to 0.051 AP25 versus 0.415, as does YOLOv8n fine-tuned identically despite 1,100 times the compute, while the deployed int8 detector reaches 0.694 AP25 in-domain and 0.444 on this split. Temporal-shift modules lower accuracy, so the deployed detector is stateless. Silent int8 calibration failures are documented; min-max calibration with propagated caches matches float within 0.008 AP. On a Raspberry Pi Zero 2W the pipeline runs at 31.85 FPS with 0.462 AP25 and 0.714 recall over 1,000 real-world UAV videos, versus 1.95 FPS and 0.172 AP25 for YOLOv8n. A 57-second field sequence shows auto-lock at 1.3 s, 97.9% lock recall, and recovery from all nine occlusions with zero false re-locks.

Figures

Figures reproduced from arXiv: 2607.16282 by Akbar Anbar Jafari, Cagri Ozcinar, Gholamreza Anbarjafari.

Figure 1
Figure 1. Figure 1: The EMTS-Det pipeline. Solid arrows carry per-frame data; dashed arrows carry the ego-motion transform [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The input representation on real drone footage (frame from the field sequence of Sec. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Stage-C architecture (deployed, stateless form). All four heads are single 1 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Precision–recall of the deployed stateless model on the held-out bench at IoU [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Recall at the operating point (τs=0.3, IoU ≥ 0.25) by ground-truth height. Performance is strong throughout the follow-me range (≥32 px) and degrades gracefully toward the ≤12-px extreme, where a person spans a single output cell. caches — loses 58% of AP25 relative to float; more calibration data does not help (rows 1–2), because the failure is the estimator, not the sample size: each additional batch pul… view at source ↗
Figure 6
Figure 6. Figure 6: Person-height distributions at 256-wide working resolution. VisDrone-DET concentrates mass at 4–8 px; [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: False positives per frame on 4,200 person-free frames from held-out videos, versus detection score threshold, with and without Stage-E verification (log scale) [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Stage-E verifier ROC on its held-out validation split ( [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Frame-by-frame comparison over the full 57-second field sequence. Bottom lane: frames on which [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: EMTS-Det (left column; green box = verified lock, yellow crosshair = tracked target) versus YOLOv8n (right column; red boxes) on the same frames of the field sequence, at t=23, 30, and 50 s. In the top two rows the target is distant (16 px tracked height) and inside one of YOLOv8n’s multi-second dropout gaps ( [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

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 · 5 linked inside Pith

  1. [1]

    Towards large-scale small object detection: Survey and benchmarks,

    G. Cheng, X. Yuan, X. Yao, K. Yan, Q. Zeng, X. Xie, and J. Han, “Towards large-scale small object detection: Survey and benchmarks,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 11, pp. 13 467–13 488, 2023

    2023

  2. [2]

    Ultralytics YOLOv8,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics YOLOv8,” https://github.com/ultralytics/ultralytics, 2023

    2023

  3. [3]

    SSD: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “SSD: Single shot multibox detector,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2016, pp. 21–37

    2016

  4. [4]

    MobileNets: Efficient convolutional neural networks for mobile vision applications,

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “MobileNets: Efficient convolutional neural networks for mobile vision applications,”arXiv preprint arXiv:1704.04861, 2017

    Pith/arXiv arXiv 2017

  5. [5]

    Visual perception of biological motion and a model for its analysis,

    G. Johansson, “Visual perception of biological motion and a model for its analysis,”Percept. Psychophys., vol. 14, no. 2, pp. 201–211, 1973. 18 Moving Like a HumanG. Anbarjafari

    1973

  6. [6]

    An iterative image registration technique with an application to stereo vision,

    B. D. Lucas and T. Kanade, “An iterative image registration technique with an application to stereo vision,” in Proc. Int. Joint Conf. Artif. Intell. (IJCAI), 1981, pp. 674–679

    1981

  7. [7]

    Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,

    M. A. Fischler and R. C. Bolles, “Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,”Commun. ACM, vol. 24, no. 6, pp. 381–395, 1981

    1981

  8. [8]

    An FFT-based technique for translation, rotation, and scale-invariant image registration,

    B. S. Reddy and B. N. Chatterji, “An FFT-based technique for translation, rotation, and scale-invariant image registration,”IEEE Trans. Image Process., vol. 5, no. 8, pp. 1266–1271, 1996

    1996

  9. [9]

    Objects as points,

    X. Zhou, D. Wang, and P. Krähenbühl, “Objects as points,”arXiv preprint arXiv:1904.07850, 2019

    Pith/arXiv arXiv 1904

  10. [10]

    TSM: Temporal shift module for efficient video understanding,

    J. Lin, C. Gan, and S. Han, “TSM: Temporal shift module for efficient video understanding,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 7083–7093

    2019

  11. [11]

    A new approach to linear filtering and prediction problems,

    R. E. Kalman, “A new approach to linear filtering and prediction problems,”J. Basic Eng., vol. 82, no. 1, pp. 35–45, 1960

    1960

  12. [12]

    Detection and tracking meet drones challenge,

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, and H. Ling, “Detection and tracking meet drones challenge,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 11, pp. 7380–7399, 2022

    2022

  13. [13]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 2704–2713

    2018

  14. [14]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 779–788

    2016

  15. [15]

    MobileNetV2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted residuals and linear bottlenecks,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 4510–4520

    2018

  16. [16]

    PP-PicoDet: A better real-time object detector on mobile devices,

    G. Yu, Q. Chang, W. Lv, C. Xu, C. Cui, W. Ji, Q. Dang, K. Deng, G. Wang, Y . Du, B. Lai, Q. Liu, X. Hu, D. Yu, and Y . Ma, “PP-PicoDet: A better real-time object detector on mobile devices,”arXiv preprint arXiv:2111.00902, 2021

    Pith/arXiv arXiv 2021

  17. [17]

    CornerNet: Detecting objects as paired keypoints,

    H. Law and J. Deng, “CornerNet: Detecting objects as paired keypoints,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2018, pp. 734–750

    2018

  18. [18]

    FCOS: Fully convolutional one-stage object detection,

    Z. Tian, C. Shen, H. Chen, and T. He, “FCOS: Fully convolutional one-stage object detection,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 9627–9636

    2019

  19. [19]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” inProc. IEEE Int. Conf. Comput. Vis. (ICCV), 2017, pp. 2980–2988

    2017

  20. [20]

    Temporal segment networks: Towards good practices for deep action recognition,

    L. Wang, Y . Xiong, Z. Wang, Y . Qiao, D. Lin, X. Tang, and L. Van Gool, “Temporal segment networks: Towards good practices for deep action recognition,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2016, pp. 20–36

    2016

  21. [21]

    MoViNets: Mobile video networks for efficient video recognition,

    D. Kondratyuk, L. Yuan, Y . Li, L. Zhang, M. Tan, M. Brown, and B. Gong, “MoViNets: Mobile video networks for efficient video recognition,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 16 020–16 030

    2021

  22. [22]

    New trends on moving object detection in video images captured by a moving camera: A survey,

    M. Yazdi and T. Bouwmans, “New trends on moving object detection in video images captured by a moving camera: A survey,”Comput. Sci. Rev., vol. 28, pp. 157–177, 2018

    2018

  23. [23]

    Simple online and realtime tracking,

    A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” inProc. IEEE Int. Conf. Image Process. (ICIP), 2016, pp. 3464–3468

    2016

  24. [24]

    Simple online and realtime tracking with a deep association metric,

    N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime tracking with a deep association metric,” in Proc. IEEE Int. Conf. Image Process. (ICIP), 2017, pp. 3645–3649

    2017

  25. [25]

    ByteTrack: Multi-object tracking by associating every detection box,

    Y . Zhang, P. Sun, Y . Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, and X. Wang, “ByteTrack: Multi-object tracking by associating every detection box,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2022, pp. 1–21

    2022

  26. [26]

    FairMOT: On the fairness of detection and re-identification in multiple object tracking,

    Y . Zhang, C. Wang, X. Wang, W. Zeng, and W. Liu, “FairMOT: On the fairness of detection and re-identification in multiple object tracking,”Int. J. Comput. Vis., vol. 129, pp. 3069–3087, 2021

    2021

  27. [27]

    HOTA: A higher order metric for evaluating multi-object tracking,

    J. Luiten, A. Ošep, P. Dendorfer, P. Torr, A. Geiger, L. Leal-Taixé, and B. Leibe, “HOTA: A higher order metric for evaluating multi-object tracking,”Int. J. Comput. Vis., vol. 129, pp. 548–578, 2021

    2021

  28. [28]

    Learning from synthetic humans,

    G. Varol, J. Romero, X. Martin, N. Mahmood, M. J. Black, I. Laptev, and C. Schmid, “Learning from synthetic humans,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 109–117

    2017

  29. [29]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), 2017, pp. 23–30. 19 Moving Like a HumanG. Anbarjafari

    2017

  30. [30]

    Quantizing deep convolutional networks for efficient inference: A whitepaper,

    R. Krishnamoorthi, “Quantizing deep convolutional networks for efficient inference: A whitepaper,”arXiv preprint arXiv:1806.08342, 2018

    Pith/arXiv arXiv 2018

  31. [31]

    A white paper on neural network quantization,

    M. Nagel, M. Fournarakis, R. A. Amjad, Y . Bondarenko, M. van Baalen, and T. Blankevoort, “A white paper on neural network quantization,”arXiv preprint arXiv:2106.08295, 2021

    Pith/arXiv arXiv 2021

  32. [32]

    Learned step size quantization,

    S. K. Esser, J. L. McKinstry, D. Bablani, R. Appuswamy, and D. S. Modha, “Learned step size quantization,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2020

    2020

  33. [33]

    Good features to track,

    J. Shi and C. Tomasi, “Good features to track,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 1994, pp. 593–600

    1994

  34. [34]

    Supervised contrastive learning,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, pp. 18 661–18 673

    2020

  35. [35]

    Two-frame motion estimation based on polynomial expansion,

    G. Farnebäck, “Two-frame motion estimation based on polynomial expansion,” inProc. Scandinavian Conf. Image Anal. (SCIA), 2003, pp. 363–370

    2003

  36. [36]

    LiteRT: Google’s on-device runtime (formerly TensorFlow Lite),

    Google AI Edge, “LiteRT: Google’s on-device runtime (formerly TensorFlow Lite),” https://ai.google.dev/edge/ litert, 2025

    2025

  37. [37]

    Blender — a 3D modelling and rendering package,

    Blender Online Community, “Blender — a 3D modelling and rendering package,” https://www.blender.org, 2024

    2024

  38. [38]

    Histograms of oriented gradients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2005, pp. 886–893. 20

    2005