Learning to Understand Body Language from Flight through Robust 3D Avatar Placing

Alina Marcu; Cristina Lazar; Dragos Costea; Marius Leordeanu

REVIEW 2 major objections 4 minor 32 references

Learning to Understand Body Language from Flight through Robust 3D Avatar Placing

T0 review · 2 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper claims that drone-distance body-language intent can be learned by placing 3D avatars into real footage, anchored to the scene by affine weights that stay valid as the camera moves.

desk verdict A useful dataset and a mostly sound training study, but the metric-placement claim is under-validated and the evaluation protocol inflates the headline numbers. read the letter →

arxiv 2607.27865 v1 pith:DUUBBPXW submitted 2026-07-30 cs.CV

Dragos Costea , Alina Marcu , Cristina Lazar , Marius Leordeanu This is my paper

classification cs.CV

keywords bodylanguageunderstandingaerialvideo3Davatarplacementintentrecognitionsyntheticdatamonoculardepthkeypointsequencesdronescenes

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Long-range body-language understanding is starved of labeled data: real actors are hard to label between 50 and 100 meters, and fully synthetic scenes lose the statistics of real aerial video. The paper's middle path is to keep the real scene and place virtual people into it. It introduces a tracker that predicts the ground point of a placed person as a fixed affine combination of static scene anchors lifted through monocular depth, with weights provably invariant to camera rotation and translation, plus a fitted ground rotation for coherent re-rendering. This yields Drones2BodyLanguage, 3,580 detection-verified keypoint sequences at 33-98 meters across ten intents and three motion sources. Training twelve skeleton-architecture baselines on placed data raises mean intent accuracy from near-chance zero-shot levels to 0.74/0.67/0.58 for real, retargeted, and generated motion, and mixed-source models transfer to two new real scenes.

What carries the argument

Frame-invariant affine anchor combination (Property 1): the placement point p is written as a sum of weights w_k times static anchors a_k lifted to 3D via monocular depth, with the constraint that the weights sum to one. Because rigid transforms act linearly and the affine constraint makes the translation factor out, the same weight vector predicts p at every frame from newly observed anchors. Around this sits the full pipeline: semantic gating of corners into anchor and placeable masks, exponential moving average on depth, a regularized least-squares fit selecting near-uniform weights, and an SVD-fitted ground rotation that re-renders the avatar frontally to the scene rather than to the swe

What would settle it

Take a scene with a surveyed static ground point, run the placement tracker, and in parallel deliberately add a per-frame constant offset to the depth map or use a depth estimator known to have per-frame shift; if the predicted affine-combination point drifts by more than about half a person-width over the 96-frame span, the metric-depth assumption is violated. A simpler check is to compare the tracker's depth output against a laser rangefinder on a static point across a panning sequence at 60-100 meters.

Watch

Extended reading notes

Core claim

The paper's core claim is that realistic, metrically correct avatar placement in unmodified drone footage is achievable from noisy monocular depth, and the resulting dataset transfers to real intent classification. The enabling observation is Property 1: a static scene point expressed as an affine combination of static anchors, with weights summing to one, keeps the same weights under any rigid camera transform. The tracker therefore predicts the placement point every frame from constant weights rather than tracking appearance, survives textureless ground, and outputs the metric depth needed for avatar scale. An SVD fit over the anchor constellation supplies the ground-plane rotation, keepin

Load-bearing premise

The load-bearing premise is that the streaming monocular depth map is metric up to a single global scale factor, because only then do lifted anchors stay rigidly related across frames and the constant affine weights keep predicting the same ground point.

Editorial extensions

If this is right

Training data for skeleton-based intent recognition can be manufactured at scale for the 50-100 meter range, where real labeling is impractical.
Scene-disjoint and motion-disjoint splits show that gains come from the motion itself, not from scene or appearance leakage.
Mixing real cut-outs with retargeted real motion is a better recipe than either alone, and a synthetic-to-real finetune chain beats direct training.
The placement tracker supplies metrically correct depth and ground rotation in closed form, not just 2D trajectories, which is exactly what compositing needs.
Detector/classifier distance bottlenecks decouple: accuracy given detection stays flat at 60-130 meters, so detection, not classification, is the range-limiting factor.

Reading between the lines

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

The affine-anchor trick could attach other virtual objects to textureless ground points in aerial video, not just human avatars, as long as the local scene is rigid and depth is roughly metric.
The ablation showing rotation costs fidelity but not learnability suggests pose-estimation benchmarks should vary yaw explicitly; models trained on yaw-augmented data may transfer better to oblique drone viewpoints.
The three-source comparison isolates a motion gap: retargeted real motion under synthetic appearance transfers far better than fully generated motion, so improving generated motion realism may matter more than improving renderer realism.
A direct test at the corpus's own 33-98 meter band, rather than the closer 3-55 meter validation scenes, would decide whether the reported transfer extends to the intended long-range regime.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The thing to know: Drones2BodyLanguage is a genuinely useful resource, and the main empirical claim — that training on placed avatars lifts intent accuracy across twelve architectures — holds up. But the fancier geometric claim, metrically correct placement over hundreds of frames, is weaker than advertised. Monocular depth with an additive shift or spatially varying bias breaks Property 1, and the real validation scenes are at shorter range than the corpus band.

What's new and good: the dataset itself, 3,580 detection-verified 80-frame sequences with three motion sources and ten intents at 33–98m, is the sort of thing the field needs. The placement pipeline is clever and mostly honest: semantically gated corners lifted to 3D, affine weights that are exactly rigid-invariant if the depth is similarity-consistent, and an SVD-fit ground rotation. The ablation showing rotation has essentially no effect on learnability is refreshing — it keeps the method's contribution honest. And the training study is broad and consistent: every architecture improves, real > realsynth > kimodo in every regime, and the real-scene validation (Countryside, Resort) confirms transfer, even if those scenes are closer than the corpus band.

The soft spots, in proportion:

1. The metric-depth assumption is the load-bearing one. Property 1 is fine as affine geometry, but it only applies if the lifted anchors are related across frames by a similarity transform. Depth Anything 3 is not guaranteed to be metric up to a global scale — per-frame shifts and spatially varying biases are common. The paper explicitly handles scale ambiguity but not shift/drift. The 96-frame tracking span may still work because the EMA and anchor replacement are robust, but the paper never measures placement drift against ground truth, so the metric scale/position claim is unsupported at the stated distances.

2. The evaluation protocol has a real flaw: the test split doubles as validation for stopping epoch (Sec 5.3), and the reported numbers are best-of-epoch. That inflates accuracy. It doesn't kill the comparison — placement vs no-placement is still likely to hold — but the absolute numbers should be read as optimistic.

3. The real validation is at 3–55m, not the 33–98m corpus range. So long-range transfer is not directly measured. Detection degrades with distance, but classification-given-detection may not; still, the claim "at distance" is only partially validated.

4. No code or dataset released. For a dataset paper, that's a major missing artifact.

None of this undermines the central direction. The training gains are broad, consistent, and externally probed. The authors have done solid engineering and the paper is clearly written.

My take: this deserves a serious referee. The right outcome is a conditional accept — the authors should release the data/code, fix or acknowledge the depth-model limitation, and either redo the epoch selection or report non-best numbers. I'd bring it to a reading group working on synthetic data or human intent recognition.

Referee Report

2 major / 4 minor

Summary. The paper introduces Drones2BodyLanguage, a dataset of 3,580 detection-verified 80-frame skeleton sequences at 33–98m, covering ten communicative intents and three motion sources (real chroma cut-outs, retargeted SMPL-X realsynth, and kimodo-generated motions), composited into eight real 4K UAV scenes. The enabling method is a 3D placement tracker that lifts Lucas–Kanade anchors through streaming monocular depth, predicts the placement point as an affine combination of anchors using weights invariant to rigid transforms (Property 1), and estimates the ground rotation by an SVD fit. The empirical core is a study across twelve skeleton-based architectures on scene- and motion-disjoint splits, showing that training on placed data raises mean intent accuracy from zero-shot levels of roughly 0.48/0.32/0.20 to 0.74/0.67/0.58 for real/realsynth/kimodo, with additional transfer results on two unseen real scenes at 3–55m. The paper is transparent about its limitations and includes extensive per-model and per-distance tables in the appendix.

Significance. If the geometric and empirical claims hold, the dataset would be a valuable resource: it addresses a real data bottleneck for long-range aerial human-intent recognition, and the compositing approach preserves real scene statistics while providing labels, distances, and controlled motion sources. The empirical contribution is substantial — consistent gains across twelve architectures, three motion sources, leakage-free splits, and two fresh real scenes, with full per-model results in Table B1. The paper is also honest about known failure modes (anchor drift, depth noise, the close-range real validation set) and explicitly discloses that the test split doubles as validation for epoch selection. Property 1 is proved correctly for rigid and similarity transforms. The main load-bearing weakness is that the actual pipeline obtains anchors from monocular depth, and the manuscript does not validate the metric/rigidity assumption on which the 'metrically correct placement maintained over hundreds of frames' claim rests. This is a correctness-risk concern for the dataset's geometric grounding, but it is not a circularity problem, and the downstream learning gains are empirically robust enoug

major comments (2)

[§3.1, Eq. (1)–(2), Appendix A] Property 1 is mathematically correct for points expressed in a common world frame under rigid or similarity transforms. However, the anchors in the actual pipeline are obtained by inverse-projecting Lucas–Kanade tracks through streaming monocular depth (Depth Anything 3), and the manuscript nowhere verifies that the resulting lifted coordinates are related across frames by a similarity transform. Appendix A explicitly addresses only a global scale ambiguity; it does not address a per-frame additive shift or a spatially varying depth bias, both common in monocular depth estimators. If the true depth is z_true = α z_pred + β_t, the lifted anchor coordinates are not a rigid/similarity transform of the true world points: the additive term is multiplied by the inverse depth and changes with each anchor's position. The affine constraint Σw_k = 1 then cannot cancel the bias, so the predicted pl
[§5.3, Table 2] The text states that 'different stop epochs were tested' and 'metrics are window-level, the test split doubling as validation.' This means the reported 'best' accuracies in Table 2 are selected on the test set, which is a form of test-set model selection. This contaminates the precise headline margins (e.g., 0.864 best on real) and the comparison between regimes, since the stop epoch is chosen per regime/model on the same test clips. The ordinal conclusion — placement training beats zero-shot, and finetuning beats from-scratch on 34 of 36 model–source pairs — is plausible and supported by Table B1, but the exact numbers are optimistic. Please either hold out a validation split for epoch selection and report test accuracy only at that epoch, or report the mean/standard deviation across epochs.

minor comments (4)

[§5.5, Table 4] The two real validation scenes are at 3–30m and 5–55m, whereas the corpus and the paper's stated operating band are 33–98m (and the zero-shot distance analysis covers 60–130m). Thus new-scene generalization and long-distance classification are each demonstrated, but not simultaneously. Please state this explicitly and, if feasible, add a real capture at 60–100m to validate the long-range transfer claim on un-composited footage.
[§3.1] The tracking-span comparison in §3.2 ('median raw corner track survives ~35 frames, anchored prediction sustains the full ~96-frame span') is reported without a methodology or a table. Please specify how these spans are measured (e.g., number of frames until the predicted point leaves a tolerance band around the ground-truth projection).
[§3.1 / Appendix A] Equation (1) uses several heuristic parameters (K=11, M=10, λ=0.5, the depth band ratio ρ, the EMA smoothing factor). A sensitivity analysis over these parameters would strengthen the claim that the method is robust; currently only λ is described as being adapted.
[General] The phrase 'metrically correct' in the abstract and contributions should be qualified to 'metrically consistent under the monocular-depth similarity assumption' until the validation requested above is performed. The current wording overstates what is verified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the placement invariance is a theorem, and the intent-learning gains are validated on un-composited real footage.

full rationale

The paper's central geometric claim, Property 1 (Sec. 3.1, Appendix A), is a mathematical statement: if a static point is expressed as an affine combination of static anchors with weights summing to one, the same weights reproduce the point under any rigid or similarity coordinate change. This is a proof, not a fitted prediction, and Eq. (2) simply evaluates a fixed affine combination of currently observed anchors. The weights are estimated once over an initialization window and then held constant; future frames provide new anchor observations. There is no fitted parameter renamed as a prediction, and no quantity is defined in terms of the target quantity. The only self-referential element is the use of the authors' prior benchmark [4] to define the ten-intent taxonomy and the twelve baseline architectures. That is task scaffolding, not a load-bearing derivation: the paper's main transfer claims are tested on un-composited real Countryside and Resort footage (Sec. 5.5), which is external to the placement pipeline. The skeptical concern about monocular depth consistency (Appendix A handles only global scale ambiguity, not per-frame shift or drift) is a correctness assumption about Depth Anything 3's metric output, not a circular step: Property 1 is conditional on the rigid/similarity premise, and the paper does not claim to have proven that premise empirically. Thus the derivation chain is self-contained, and any weakness is an unvalidated assumption, not circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central engineering claim is the frame-invariant anchor tracker: a static ground point is represented as an affine combination of K static anchors, with weights fixed after 10 frames and immune to rigid camera motion. This rests on world-static anchors, monocular depth consistent under similarity transforms, and accurate semantic masks. The method introduces hand-set hyperparameters (λ, K, M, depth band ρ, EMA) whose values are partly undisclosed. No invented physical entities are introduced.

free parameters (5)

Tikhonov regularization λ = 0.5 (doubled if max_k |w_k| > 0.5)
Hand-chosen to stabilize the least-squares weight estimation; not fitted to intent accuracy.
Number of anchors K = 11
Hand-chosen; affects conditioning and local rigidity of the anchor constellation.
Initialization window M = 10 frames
Hand-chosen; weights are estimated over this window.
Depth band ratio ρ = not stated
Used to select anchors with depth near the placement point; chosen by hand, no value given.
EMA smoothing factor for depth = not stated
Exponential moving average stabilizes per-query depth; smoothing factor not disclosed.

assumptions (6)

domain assumption Anchors and placement points are world-static across the tracking window.
Property 1 depends on p and a_k being static; LK tracking drift is handled by replacement, but any residual static violation biases prediction.
domain assumption Monocular depth from [15] is metric up to a global scale factor.
Needed for inverse projection to yield coordinates rigidly consistent across frames; shift/scale ambiguity breaks Property 1.
domain assumption Semantic masks (SAM3) accurately separate placeable ground from anchorable structures.
Placement points sampled on the mask must lie on walkable ground; errors bias scale and position.
domain assumption Ground patch is locally rigid/planar so a single SVD rotation describes the anchor constellation.
Sec 3.2 fits one R_t for all anchors; non-rigid scene motion breaks the fit.
domain assumption The benchmark protocol of [4] (60 frames forecast 30 + classify 10 intents) is a valid proxy for communicative intent.
Evaluation uses this protocol; the intents are defined by the authors' prior paper.
domain assumption YOLO11x-pose detection is reliable enough to verify placements and select real-scene clips.
Placement harvesting and real validation clips are both filtered by YOLO11x-pose detection; detector failures are excluded.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Understand Body Language from Flight through Robust 3D Avatar Placing." pith.science (2026). https://pith.science/paper/DUUBBPXW

@misc{pith2026260727865,
  author       = {Pith},
  title        = {Pith review of: Learning to Understand Body Language from Flight through Robust 3D Avatar Placing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUUBBPXW}},
  note         = {Machine review of arXiv:2607.27865}
}

read the original abstract

Perceiving human motion and intent at long range is a prerequisite for socially intelligent aerial robots, yet the data to learn it barely exists. We introduce Drones2BodyLanguage, a dataset grounding human motion in real UAV footage: avatars manifesting ten communicative intents are placed into unmodified 4K drone scenes with metrically correct position, scale and orientation, maintained over hundreds of frames of camera motion. Enabling it is a lightweight geometric world model of the local scene - semantically selected anchors lifted to 3D through streaming monocular depth - in which a placement point is predicted as an affine anchor combination with provably rigid-invariant weights, and re-rendered under an SVD-fitted ground rotation. Across twelve architectures on scene- and motion-disjoint splits, training on placed data lifts mean intent accuracy by a wide margin for real, retargeted and generated motion alike, with gains confirmed on two in-the-wild scenes.

Figures

Figures reproduced from arXiv: 2607.27865 by the authors.

**Figure 1.** Overview: real and synthetic motions and personas are composited into real 4K drone scenes via SAM3 segmentation, metric depth and our 3D placement tracker; the real/realsynth/kimodo placements drive the intent study at distance answer is a streaming geometric world model: semantically gated corners on world-static structure are lifted to 3D through noisy monocular depth, and the placement point is an affine combina… view at source ↗

**Figure 2.** Samples from the two real validation scenes, all ten intents: person-centred crops from the Countryside (C, first three) and Resort (R) from 4K footage [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 1 canonical work pages

[1]

Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Surís, D., Ryali, C., et al.: SAM 3: Segment anything with concepts. Tech. rep., Meta Superintelli- gence Labs (Nov 2025),https://ai.meta.com/research/publications/sam-3- segment-anything-with-concepts/

2025
[2]

In: ICCV

Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topology refinement graph convolution for skeleton-based action recognition. In: ICCV. pp. 13359–13368 (2021)

2021
[3]

In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

Cong, W., Zhang, J., Niu, L., Liu, L., Ling, Z., Li, W., Zhang, L.: DoveNet: Deep image harmonization via domain verification. In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 8394–8403 (2020)

2020
[4]

Costea, D., Marcu, A., Lazar, C., Leordeanu, M.: Non-verbal real-time human-ai interaction in constrained robotic environments (2026),https://arxiv.org/abs/ 2603.01804

arXiv 2026
[5]

arXiv preprint arXiv:2404.19759 (2024)

Dai, W., Chen, L.H., Wang, J., Liu, J., Dai, B., Tang, Y.: Motionlcm: Real- time controllable motion generation via latent consistency model. arXiv preprint arXiv:2404.19759 (2024)

arXiv 2024
[6]

In: IEEE/CVF International Conference on Computer Vision (ICCV)

Doersch, C., Yang, Y., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y., Carreira, J., Zisserman, A.: TAPIR: Tracking any point with per-frame initialization and temporal refinement. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10061–10072 (2023)

2023
[7]

In: IEEE International Conference on Computer Vision (ICCV)

Dwibedi,D.,Misra,I.,Hebert,M.:Cut,pasteandlearn:Surprisinglyeasysynthesis for instance detection. In: IEEE International Conference on Computer Vision (ICCV). pp. 1301–1310 (2017)

2017
[8]

com / ultralytics/ultralytics

Jocher, G., Qiu, J.: Ultralytics YOLO11 (2024),https : / / github . com / ultralytics/ultralytics

2024

Show all 32 references

[9]

Acta Crystallographica Section A32(5), 922–923 (1976)

Kabsch, W.: A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A32(5), 922–923 (1976)

1976
[10]

In: European conference on computer vision

Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. In: European conference on computer vision. pp. 18–35. Springer (2024)

2024
[11]

International Journal of Con- trol, Automation and Systems (2025).https://doi.org/10.1007/s12555-025- 0012-y

Kim, J., et al.: A human-following drone providing gesture recognition to control IoT devices based on 3D body-landmark detection. International Journal of Con- trol, Automation and Systems (2025).https://doi.org/10.1007/s12555-025- 0012-y

2025 doi
[12]

International Journal of Computer Vision81(2), 155–166 (2009)

Lepetit, V., Moreno-Noguer, F., Fua, P.: EPnP: An accurateO(n)solution to the PnP problem. International Journal of Computer Vision81(2), 155–166 (2009)

2009
[13]

In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) Learning Body Language from Flight 9

Li, Z., Tucker, R., Cole, F., Wang, Q., Jin, L., Ye, V., Kanazawa, A., Holynski, A., Snavely, N.: MegaSaM: Accurate, fast and robust structure and motion from ca- sual dynamic videos. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) Learning Bod...

2025
[14]

Lin, C.H., Yumer, E., Wang, O., Shechtman, E., Lucey, S.: ST-GAN: Spatial trans- formergenerativeadversarialnetworksforimagecompositing.In:IEEEConference on Computer Vision and Pattern Recognition (CVPR). pp. 9455–9464 (2018)

2018
[15]

arXiv preprint arXiv:2511.10647 (2025)

Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025)

2025 arXiv
[16]

IEEE Trans- actions on Pattern Analysis and Machine Intelligence42(10), 2684–2701 (2020)

Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L.Y., Kot, A.C.: NTU RGB+D 120: A large-scale benchmark for 3D human activity understanding. IEEE Trans- actions on Pattern Analysis and Machine Intelligence42(10), 2684–2701 (2020)

2020
[17]

arXiv preprint arXiv:2107.01889 (2021)

Liu, L., Liu, Z., Zhang, B., Li, J., Niu, L., Liu, Q., Zhang, L.: OPA: Object place- ment assessment dataset. arXiv preprint arXiv:2107.01889 (2021)

2021 arXiv
[18]

In: Proceedings of the 7th International Joint Conference on Artificial Intelligence (IJCAI)

Lucas, B.D., Kanade, T.: An iterative image registration technique with an appli- cation to stereo vision. In: Proceedings of the 7th International Joint Conference on Artificial Intelligence (IJCAI). pp. 674–679 (1981)

1981
[19]

In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10975–10985 (2019)

2019
[20]

Sensors21(6), 2180 (2021)

Perera, A.G., Law, Y.W., Ogunwa, T.T., Chahl, J.: Real-time human detection and gesture recognition for on-board UAV rescue. Sensors21(6), 2180 (2021)

2021
[21]

arXiv preprint arXiv:2603.15546 (2026), nVIDIA Toronto AI Lab technical report

Rempe, D., Petrovich, M., Yuan, Y., Zhang, H., Peng, X.B., Jiang, Y., Wang, T., Iqbal, U., Minor, D., de Ruyter, M., Li, J., Tessler, C., Lim, E., Jeong, E., Wu, S., Hassani, E., Huang, M., Yu, J.B., Chung, C., Song, L., Dionne, O., Kautz, J., Yuen, S., Fidler, S.: Kimodo: Sca...

2026
[22]

In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Shi, J., Tomasi, C.: Good features to track. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 593–600 (1994)

1994
[23]

In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Tripathi, S., Chandra, S., Agrawal, A., Tyagi, A., Rehg, J.M., Chari, V.: Learning to generate synthetic data via compositing. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 461–470 (2019)

2019
[24]

IEEE Transactions on Pattern Analysis and Machine Intelligence 13(4), 376–380 (1991)

Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence 13(4), 376–380 (1991)

1991
[25]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Varol, G., Romero, J., Martin, X., Mahmood, N., Black, M.J., Laptev, I., Schmid, C.: Learning from synthetic humans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 109–117 (2017)

2017
[26]

In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: VGGT: Visual geometry grounded transformer. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

2025
[27]

In: IEEE/CVF International Conference on Computer Vision (ICCV) (2025)

Xiao, Y., Wang, J., Xue, N., Karaev, N., Makarov, Y., Kang, B., Zhu, X., Bao, H., Shen, Y., Zhou, X.: SpatialTrackerV2: 3D point tracking made easy. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2025)

2025
[28]

In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

Xiao, Y., Wang, Q., Zhang, S., Xue, N., Peng, S., Shen, Y., Zhou, X.: Spatial- Tracker: Tracking any 2D pixels in 3D space. In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 20406–20417 (2024)

2024
[29]

In: AAAI

Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: AAAI. pp. 7444–7452 (2018)

2018
[30]

In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 10 D

Zhang, B., Ke, L., Harley, A.W., Fragkiadaki, K.: TAPIP3D: Tracking any point in persistent 3D geometry. In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 10 D. Costea et al

2025
[31]

In: European Conference on Computer Vision (ECCV)

Zhou,S.,Liu,L.,Niu,L.,Zhang,L.:Learningobjectplacementviadual-pathgraph completion. In: European Conference on Computer Vision (ECCV). pp. 373–389 (2022)

2022
[32]

static anchor

Zhu, S., Lin, Z., Cohen, S., Kuen, J., Zhang, Z., Chen, C.: TopNet: Transformer- based object placement network for image compositing. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1838–1847 (2023) Learning Body Language from Flight 11 Appendix...

2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Surís, D., Ryali, C., et al.: SAM 3: Segment anything with concepts. Tech. rep., Meta Superintelli- gence Labs (Nov 2025),https://ai.meta.com/research/publications/sam-3- segment-anything-with-concepts/

2025

[2] [2]

In: ICCV

Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topology refinement graph convolution for skeleton-based action recognition. In: ICCV. pp. 13359–13368 (2021)

2021

[3] [3]

In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

Cong, W., Zhang, J., Niu, L., Liu, L., Ling, Z., Li, W., Zhang, L.: DoveNet: Deep image harmonization via domain verification. In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 8394–8403 (2020)

2020

[4] [4]

Costea, D., Marcu, A., Lazar, C., Leordeanu, M.: Non-verbal real-time human-ai interaction in constrained robotic environments (2026),https://arxiv.org/abs/ 2603.01804

arXiv 2026

[5] [5]

arXiv preprint arXiv:2404.19759 (2024)

Dai, W., Chen, L.H., Wang, J., Liu, J., Dai, B., Tang, Y.: Motionlcm: Real- time controllable motion generation via latent consistency model. arXiv preprint arXiv:2404.19759 (2024)

arXiv 2024

[6] [6]

In: IEEE/CVF International Conference on Computer Vision (ICCV)

Doersch, C., Yang, Y., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y., Carreira, J., Zisserman, A.: TAPIR: Tracking any point with per-frame initialization and temporal refinement. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10061–10072 (2023)

2023

[7] [7]

In: IEEE International Conference on Computer Vision (ICCV)

Dwibedi,D.,Misra,I.,Hebert,M.:Cut,pasteandlearn:Surprisinglyeasysynthesis for instance detection. In: IEEE International Conference on Computer Vision (ICCV). pp. 1301–1310 (2017)

2017

[8] [8]

com / ultralytics/ultralytics

Jocher, G., Qiu, J.: Ultralytics YOLO11 (2024),https : / / github . com / ultralytics/ultralytics

2024