PoseConvSpot: Overcoming Domain Shifts in Precise Event Spotting for Sports Videos via Pose-Only Spatio-Temporal Regression

ACM MMSports '26 (ACM Multimedia 2026 Workshop)
Noâm Boussouf1,2,*, Adrien Maglo3, Mohamed Chaouch3, Paul Pham Dang1,2, Hideo Saito1
1Keio University, Japan   2École Centrale de Nantes, France   3Université Paris-Saclay, CEA-List, France
*Corresponding Author
Paper Supplementary Code (Coming Soon) Dataset (Coming Soon)
PoseConvSpot Teaser

PoseConvSpot isolates athletes by projecting 2D pose skeletons into spatio-temporal video heatmaps, mitigating visual domain shifts and outperforming state-of-the-art RGB spotting baselines by +14.1% mAP@3 in unseen arenas.

News & Updates

  • [August 2026] Project webpage launched and camera-ready paper accepted at ACM MMSports '26.
  • [November 2026] Oral presentation at ACM MMSports '26 (Rio de Janeiro, Brazil).
  • [Coming Soon] Public release of the training/inference code, dataset extraction pipeline, and pretrained checkpoints.
  • [Coming Soon] Public release of the video teaser.

Abstract

Precise Event Spotting (PES) demands frame-level accuracy for localized actions in sports analytics. Existing RGB pipelines struggle in dense multi-player scenes and are vulnerable to pixel-level domain shifts (lighting, court layouts, camera angles). We propose PoseConvSpot, a pose-only spatio-temporal regression framework tailored for high-fidelity micro-event spotting. Operating exclusively on 2D joint coordinates, our pipeline projects skeletons into spatio-temporal heatmaps that naturally isolate individual actors. Combining directional Strip Pooling, a Shared-Head GRU, and a continuous Soft Gamma Loss, PoseConvSpot isolates biomechanical events without relying on fine-grained visual textures. Evaluations on a multi-sport benchmark show that our method matches state-of-the-art RGB architectures while providing superior cross-arena generalization (+14.1% absolute mAP@3 gain on unseen courts) and peak-focused predictions without heuristic post-processing.

Method Overview

Architecture Diagram

PoseConvSpot decouples event spotting from visual appearance by relying exclusively on 2D kinematic features:

  1. Spatio-Temporal Pose Heatmaps: 2D skeletal detections are stacked over temporal sliding windows into pseudo-image heatmaps, effectively isolating individual athlete motions without visual background noise.
  2. Strip Pooling & Recurrent Head: Directional strip pooling captures long-range spatial joint trajectories, while a Shared-Head GRU models frame-by-frame temporal dynamics.
  3. Continuous Soft Gamma Loss: Mitigates temporal label imbalance and enforces smooth, confident peak generation around exact frame-level event timestamps.

Experimental Results

Evaluation on the SportsJumpMotion dataset across standard, cross-venue (LOAO), and data-constrained protocols.

Standard Benchmark (SOTA Comparison)

Standard protocol on volleyball and basketball tracklets. Top results averaged over 3 seeds.

Method Mod. mAP@1 mAP@2 mAP@3 mAP@4 F1 (@3)
E2E-Spot RGB 54.31 74.84 80.55 81.86 84.27
T-DEED200MF RGB 82.67 ±0.70 91.68 ±1.00 94.41 ±0.40 94.95 ±0.20 91.52 ±0.81
T-DEED800MF RGB 84.52 ±1.70 92.28 ±0.30 94.66 ±0.40 95.45 ±0.40 91.71 ±0.15
AdaSpot RGB 82.06 90.11 92.91 93.78 87.28
Ours (Joints) Pose 78.83 ±1.62 92.26 ±0.43 95.91 ±0.40 96.79 ±0.12 91.62 ±0.31
Ours (Limbs) Pose 79.09 ±1.06 93.20 ±0.64 96.05 ±0.66 96.80 ±0.50 92.32 ±0.18
Ours (Full) Pose 79.08 ±0.34 93.82 ±0.20 96.88 ±0.32 97.70 ±0.32 93.16 ±0.52

PoseConvSpot outperforms RGB baselines on all operational thresholds ($\delta \ge 2$) without relying on visual background cues.

Unseen Domain Generalization (LOAO)

Leave-One-Arena-Out protocol (trained on a court subset, tested on novel venues).

Method Mod. mAP@1 mAP@2 mAP@3 mAP@4
T-DEED200MF RGB 58.11 74.50 76.35 76.93
T-DEED800MF RGB 61.79 72.61 73.32 73.75
Ours (Joints) Pose 65.69 83.40 89.57 90.30
Ours (Limbs) Pose 66.31 82.83 87.95 90.85
Ours (Full J+L) Pose 64.58 81.91 90.40 91.23

+14.05% absolute gain in mAP@3 over RGB baselines, demonstrating immunity against unseen lighting and court designs.

Data-Efficiency Under Data Constraints

Performance when restricting the training set to 50% ($\text{Data}_{0.5}$) and 10% ($\text{Data}_{0.1}$) of available videos.

Data Split Method Modality mAP@1 mAP@2 mAP@3 mAP@4
$\text{Data}_{0.5}$
(50% Data) 		T-DEED200MF RGB 79.59 88.39 91.40 92.09
T-DEED800MF RGB 79.82 88.85 91.70 92.51
Ours (Joints) Pose 76.46 91.65 95.43 96.30
Ours (Limbs) Pose 75.89 91.97 95.45 96.48
Ours (Full J+L) Pose 76.31 91.71 95.08 96.09
$\text{Data}_{0.1}$
(10% Data) 		T-DEED200MF RGB 59.07 70.87 76.17 77.47
T-DEED800MF RGB 62.21 75.12 79.29 81.23
Ours (Joints) Pose 66.29 86.04 91.33 92.98
Ours (Limbs) Pose 66.03 86.59 91.81 93.63
Ours (Full J+L) Pose 67.57 87.32 91.76 93.68

With only 10% of training data, PoseConvSpot retains 91.76% mAP@3 (+12.47% over the top RGB baseline), demonstrating that kinematic filtering eliminates the need for large training datasets.

BibTeX

@inproceedings{boussouf2026poseconvspot,
  title={PoseConvSpot: Overcoming Domain Shifts in Precise Event Spotting for Sports Videos via Pose-Only Spatio-Temporal Regression},
  author={Boussouf, No{\^a}m and Maglo, Adrien and Chaouch, Mohamed and Pham Dang, Paul and Saito, Hideo},
  booktitle={Proceedings of the 9th ACM International Workshop on Multimedia Content Analysis in Sports (MMSports '26)},
  year={2026},
  doi={10.1145/3841455.3841529}
}