Research map / Detection and segmentation / Tracking
Two-stream Beats One-stream: Asymmetric Siamese Network for Efficient Visual Tracking
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 AsymTrack, an asymmetric Siamese tracker whose template branch runs once at initialization and feeds prototypes into the search branch via efficient template modulation, plus an re-parameterizable object perception enhancement module for fast visual tracking on CPU/edge devices.
Problem
Efficient visual tracking on resource-constrained platforms: one-stream trackers recompute template features every frame and use heavy transformer attention, while two-stream Siamese trackers lack template-search interaction and accuracy.
Approach
An asymmetric Siamese design that computes the template only once and unidirectionally injects template prototypes into the search branch through dynamic separable 1D convolution and prototype attention (ETM), combined with a re-parameterizable OPE module that fuses 3x3 semantic and derived local 1x1-style descriptors.
Key results
- AsymTrack-T reaches 60.8% AUC on LaSOT and 224/81/84 FPS on GPU/CPU/AGX, 6.0% AUC above HiT-Tiny.
- AsymTrack-B achieves 67.7% AO on GOT-10k, the best real-time score, and 80.0% AUC on TrackingNet.
- AsymTrack-B reaches 25.4% EAO on VOT2021 real-time testing on Jetson AGX Xavier, above other real-time efficient trackers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AsymTrack-B instead of HiT-Base | real-time tracking · GOT-10k | AO: 64.0 -> 67.7 | Helped |
| Use AsymTrack-B instead of HiT-Base | real-time tracking · LaSOT | AUC: 64.6 -> 64.7 | No clear effect |
| Use AsymTrack-T instead of HiT-Tiny | real-time tracking · LaSOT | AUC: +6.0% AUC | Helped |
| Add ETM (ablation) | AsymTrack-S · GOT-10k | AO: 60.4 -> 62.9 | Helped |
| Add ETM (ablation) | AsymTrack-S · LaSOT | AUC: 58.8 -> 61.2 | Helped |
| Add OPE (ablation) | AsymTrack-S · GOT-10k | AO: 60.4 -> 62.1 | Helped |
| Add OPE (ablation) | AsymTrack-S · LaSOT | AUC: 58.8 -> 60.7 | Helped |
| Replace DTM with vanilla MHCA (ablation) | AsymTrack-S ETM · GOT-10k | AO: 64.4 -> 64.1 | Hurt |
| Replace DTM with vanilla MHCA (ablation) | AsymTrack-S ETM · LaSOT | AUC: 62.5 -> 62.6 | No clear effect |
| Remove Prototype Attention (ablation) | AsymTrack-S ETM · GOT-10k | AO: 64.4 -> 63.9 | Hurt |
| Remove Prototype Attention (ablation) | AsymTrack-S ETM · LaSOT | AUC: 62.5 -> 62.3 | Hurt |
| Add ETM at stage 1 (ablation) | AsymTrack-S ETM · GOT-10k | AO: 64.4 -> 64.6 | No clear effect |
Limitations
- Remaining accuracy gap versus precision-oriented trackers, largest on low resolution, viewpoint change, and fast motion.
- Speed and accuracy comparisons depend on specific hardware (2080Ti, i7-9700KF, Jetson AGX Xavier).
Experimental setup
Datasets: LaSOT TrackingNet COCO2017 GOT-10k
Benchmarks: GOT-10k LaSOT TrackingNet NFS UAV123 LaSOT_ext VOT2021
Metrics: AO SR0.5 SR0.75 AUC normalized precision precision EAO FPS
Baselines: HiT-Tiny HiT-Small HiT-Base MixFormerV2 TransT OSTrack-256 ARTrack ECO LightTrack E.T.Track FEAR TCTrack
Builds on: EfficientMod STARK MixFormer HiT OSTrack
Closest papers by meaning
- STMTrack: Template-Free Visual Tracking With Space-Time Memory Networks (CVPR 2021 · similarity 0.88)
- FARTrack: Fast Autoregressive Visual Tracking with High Performance (ICLR 2026 · similarity 0.87)
- MUTrack: A Memory-Aware Unified Representation Framework for Visual Tracking (AAAI 2026 · similarity 0.87)
- LightTrack: Finding Lightweight Neural Networks for Object Tracking via One-Shot Architecture Search (CVPR 2021 · similarity 0.86)
- UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking (CVPR 2026 · similarity 0.86)
- OneTracker: Unifying Visual Object Tracking with Foundation Models and Efficient Tuning (CVPR 2024 · similarity 0.85)
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.