Research map / Video / Video-language understanding
Temporal Saliency Query Network for Efficient Video Recognition
Boyang Xia, Zhihao Wang, Wenhao Wu, Haoran Wang, Jungong Han
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 class-specific attention and classifier (CS+CS) over class-agnostic (ablation) | TSQNet VQM · ActivityNet (ResNet-101) | mAP (%): 74.0 -> 74.7 | Improved |
| use class-specific classifier together with class-specific attention (ablation) | TSQNet VQM · ActivityNet (ResNet-101) | mAP (%): 68.7 -> 74.7 | Improved |
| fuse VQM and TQM without interaction losses (ablation) | TSQNet · ActivityNet (ResNet-101) | mAP (%): 72.0 -> 74.9 (TQM), 74.6 -> 74.9 (VQM) | Improved |
| add both cross-modal interaction losses (L t->v and L v->t) (ablation) | TSQNet · ActivityNet (ResNet-101) | mAP (%): 74.9 -> 75.3 | Improved |
| add L v->t loss alone (ablation) | TQM · ActivityNet (ResNet-101) | mAP (%): 72.0 -> 72.7 | Improved |
| use Bert with top-10 object scores instead of Word2Vec (ablation) | TQM · ActivityNet (ResNet-101) | mAP (%): 71.2 -> 72.1 | Improved |
| keep only top-10 object categories instead of all (ablation) | TQM with Bert · ActivityNet (ResNet-101) | mAP (%): 71.4 -> 72.1 | Improved |
| initialize visual TSQ embedding with prototypes instead of random (ablation) | VQM · ActivityNet (ResNet-101) | mAP (%): 73.8 -> 74.7 | Improved |
| initialize textual TSQ embedding with Bert embedding instead of random (ablation) | TQM · ActivityNet (ResNet-101) | mAP (%): 71.6 -> 72.1 | Improved |
| add self-attention layer over TSQ queries (ablation) | TSQNet VQM · ActivityNet (ResNet-101) | mAP (%): 74.7 -> 74.5 | Hurt |
| use 8 heads instead of 1 in cross-attention (ablation) | TSQNet VQM · ActivityNet (ResNet-101) | mAP (%): 74.7 -> 73.7 | Hurt |
| use 2 layers instead of 1 in cross-attention (ablation) | TSQNet VQM · ActivityNet (ResNet-101) | mAP (%): 74.7 -> 73.4 | Hurt |
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.
- SCSampler — SCSampler: Sampling Salient Clips From Video for Efficient Action Recognition (ICCV 2019) · 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
- Revisiting Video Saliency: A Large-Scale Benchmark and a New Model (CVPR 2018 · similarity 0.89)
- NSNet: Non-Saliency Suppression Sampler for Efficient Video Recognition (ECCV 2022 · similarity 0.89)
- Revisiting Video Saliency Prediction in the Deep Learning Era (TPAMI 2019 · similarity 0.89)
- Follow the Saliency: Supervised Saliency for Retrieval-augmented Dense Video Captioning (CVPR 2026 · similarity 0.89)
- CaRDiff: Video Salient Object Ranking Chain of Thought Reasoning for Saliency Prediction with Diffusion (AAAI 2025 · similarity 0.88)
- A Motion-aware Spatio-temporal Graph for Video Salient Object Ranking (NeurIPS 2024 · 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.
