Research map / Video / Action recognition
SV-TAD: Native Sparse Convolutions for Efficient Temporal Action Detection
Ricardo Ignacio Pizarro Carreño, Roberto Valle, José Buenaposada, Luis M. Bergasa, Luis Baumela
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-attention module (ablation) | SV-TAD · THUMOS-14 (VideoMAEv2-B) | average mAP: 71.80 -> 72.44 | Improved |
| use SparseConv2D instead of DenseConv2D (ablation) | SV-TAD adapter · THUMOS-14 | average mAP: 68.83 -> 69.35 | Improved |
| use DenseConv1D instead of SparseConv2D (ablation) | SV-TAD adapter · THUMOS-14 | average mAP: 69.35 -> 69.60 | Improved |
| run both convs in FP32 without token selection (ablation) | SparseConv2D vs DenseConv2D · THUMOS-14 | average mAP: 69.79 vs 69.89 | No clear effect |
| add auxiliary landmark supervision (Kp) (ablation) | SV-TAD · ATTACH | average mAP: 16.28 -> 18.69 | Improved |
| add cross-attention without auxiliary tokens (ablation) | SV-TAD · ATTACH | average mAP: 17.09 -> 16.28 | Hurt |
| lower keep rate from 0.7 to 0.5 (ablation) | SV-TAD InternVideoNext-L · THUMOS-14 | average mAP: 74.13 -> 73.00 | Mixed |
| lower keep rate from 0.7 to 0.6 (ablation) | SV-TAD InternVideoNext-L · THUMOS-14 | average mAP: 74.13 -> 73.64 | Mixed |
| use EViT-style attention selector instead of default attention selector (ablation) | SV-TAD · THUMOS-14 | average mAP: 72.44 -> 72.43 | No clear effect |
| use norm-based selector instead of attention-based (ablation) | SV-TAD · THUMOS-14 | average mAP: 72.44 -> 68.06 | Hurt |
| replace AIM dense spatial adapter with SparseConv2D plus token selection (ablation) | AIM CLIP ViT-B/16 · Diving-48 | top-1 accuracy: 88.9 -> 88.7 | No clear effect |
| replace dense AdaTAD with native sparse SV-TAD | AdaTAD VMAEv2-B · THUMOS-14 | TFLOPs: 17.9 -> 10.29 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TALLFormer — TALLFormer: Temporal Action Localization with a Long-Memory Transformer (ECCV 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Scaling Action Detection: AdaTAD++ with Transformer-Enhanced Temporal-Spatial Adaptation (ICCV 2025 · similarity 0.90)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.90)
- End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames (CVPR 2024 · similarity 0.89)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.88)
- TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression (CVPR 2024 · similarity 0.88)
- Temporal Action Detection Model Compression by Progressive Block Drop (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
