Research map / Detection and segmentation / Tracking
Learning Random Noise Salient Feature Fusion Siamese Network for Low-Resolution Object Tracking (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a Siamese tracker (SiamFC-based) that fuses noise-infused feature maps and CBAM-style salient attention to improve object tracking in low-resolution scenes, evaluated on OTB100 subsets.
Problem
Siamese trackers perform sub-optimally on low-resolution (LR) object tracking, where target details are sparse and overfitting to similar frame pairs is common.
Approach
Fuse feature maps of original and noise-infused inputs (Gaussian, salt-and-pepper, Poisson, speckle, blur, sharpen) as a regularizer during training, and add block-wise CBAM salient attention on the fused features to focus on target regions in SiamFC.
Key results
- Achieves success score 0.685 and precision 0.937 on OTB100 LR-attribute subsets, above TransT (0.678/0.891) and AiATrack (0.663/0.918).
- Runs at 68 FPS, over real-time, versus 50 FPS for TransT and 38 FPS for AiATrack.
- SiamFC baseline scores 0.573 success and 0.815 precision in the same comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add random-noise feature fusion and salient attention | SiamFC · OTB100 LR challenges | success: 0.573 -> 0.685 | Helped |
| Add random-noise feature fusion and salient attention | SiamFC · OTB100 LR challenges | precision: 0.815 -> 0.937 | Helped |
| Add random-noise feature fusion and salient attention | SiamFC · OTB100 LR challenges | speed (FPS): 86 -> 68 | Hurt |
Surprising findings
- SiamFC runs faster (86 FPS) than the proposed method (68 FPS), which the paper does not discuss in depth.
Limitations
- Evaluation is limited to a small set of baselines and one benchmark (OTB100 with LR-related attributes) in a short student abstract.
- Noise types and injection layers/proportions were tuned empirically with few details given.
- Hardware and training data size are not reported.
Experimental setup
Datasets: OTB100
Benchmarks: OTB100
Metrics: success (AUC) precision FPS
Baselines: TransT AiATrack SCSAtt SiamFC
Builds on: SiamFC CBAM
Closest papers by meaning
- Deeper and Wider Siamese Networks for Real-Time Visual Tracking (CVPR 2019 · similarity 0.87)
- Learning To Fuse Asymmetric Feature Maps in Siamese Trackers (CVPR 2021 · similarity 0.87)
- SiamRPN++: Evolution of Siamese Visual Tracking With Very Deep Networks (CVPR 2019 · similarity 0.87)
- Learning To Filter: Siamese Relation Network for Robust Tracking (CVPR 2021 · similarity 0.86)
- Correlation-Aware Deep Tracking (CVPR 2022 · similarity 0.86)
- Discriminative and Robust Online Learning for Siamese Visual Tracking (AAAI 2020 · similarity 0.86)
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.