Research map / Video / Action recognition
NSNet: Non-Saliency Suppression Sampler for Efficient Video Recognition
Boyang Xia, Wenhao Wu, Haoran Wang, Rui Su, Dongliang He, Haosen Yang, Xiaoran Fan, Wanli Ouyang
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 the non-saliency suppression loss (ablation) | Video Glimpse Module · ActivityNet | mAP (%): 73.4 -> 73.8 | Improved |
| add non-saliency suppression pseudo labels (ablation) | Frame Scrutinize Module · ActivityNet | mAP (%): 68.4 -> 74.7 | Improved |
| add a non-salient class without NS supervision (ablation) | Frame Scrutinize Module · ActivityNet | mAP (%): 68.4 -> 68.6 | No clear effect |
| add a non-salient class without NS supervision (ablation) | Video Glimpse Module · ActivityNet | mAP (%): 73.4 -> 73.4 | No clear effect |
| use transformer encoder instead of 1D Conv in feature embedding (ablation) | NSNet · ActivityNet | mAP (%): 73.6 -> 75.5 | Improved |
| use transformer encoder instead of MLP in feature embedding (ablation) | NSNet · ActivityNet | mAP (%): 74.3 -> 75.5 | Improved |
| use NS objective instead of soft-label baseline+ for frame saliency (ablation) | Frame Scrutinize Module · ActivityNet | mAP (%): 73.7 -> 74.7 | Improved |
| use prototype-based instead of response-based guiding saliency score (ablation) | Frame Scrutinize Module · ActivityNet | mAP (%): 74.1 -> 74.7 | Improved |
| fuse FSM and VGM saliency instead of using FSM alone (10 frames) (ablation) | NSNet · ActivityNet | mAP (%): 74.7 -> 75.5 | Improved |
| sample 10 frames instead of 5 (ablation) | NSNet · ActivityNet | mAP (%): 74.9 -> 75.5 | Improved |
| use MobileNetv2 instead of ShuffleNetv2 as lightweight extractor (ablation) | NSNet · ActivityNet | mAP (%): 70.8 -> 75.5 | Improved |
| use EfficientNet-b0 instead of MobileNetv2 as lightweight extractor (ablation) | NSNet · ActivityNet | mAP (%): 75.5 -> 76.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AR-Net — AR-Net: Adaptive Frame Resolution for Efficient Action Recognition (ECCV 2020) · better on 1 of 1 shared result
- 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
- Temporal Saliency Query Network for Efficient Video Recognition (ECCV 2022 · similarity 0.89)
- SALISA: Saliency-Based Input Sampling for Efficient Video Object Detection (ECCV 2022 · similarity 0.86)
- Learning to Zoom: a Saliency-Based Sampling Layer for Neural Networks (ECCV 2018 · similarity 0.86)
- TASED-Net: Temporally-Aggregating Spatial Encoder-Decoder Network for Video Saliency Detection (ICCV 2019 · similarity 0.86)
- An Iterative and Cooperative Top-Down and Bottom-Up Inference Network for Salient Object Detection (CVPR 2019 · similarity 0.86)
- CaRDiff: Video Salient Object Ranking Chain of Thought Reasoning for Saliency Prediction with Diffusion (AAAI 2025 · similarity 0.84)
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.
