Research map / Detection and segmentation / Tracking
MambaLCT: Boosting Tracking via Long-term Context State Space Model
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 MambaLCT, a single-object tracker that uses a unidirectional Context Mamba module to aggregate target-change cues from the first frame to the current frame and injects them into a hierarchical-ViT attention encoder.
Problem
Visual object tracking relies on short-term context (adjacent frames or fixed clips) or the initial template alone, so long-term target variation is under-utilized; Transformers' quadratic cost limits context length.
Approach
Use a unidirectional Mamba over the temporal sequence of search-frame features, with a context token carrying hidden state across frames, so the context covers the whole video history; this context token is injected into the ViT attention with template and search tokens, while Mamba is applied only along time, not space.
Key results
- MambaLCT-256 reaches 71.8% AUC on LaSOT and 74.8% AO on GOT-10K, with 51.6% AUC on LaSOT_ext.
- MambaLCT-384 achieves 73.6% AUC on LaSOT, surpassing ODTrack-384 (73.2%), and 85.2% AUC on TrackingNet.
- Ablation: the Mamba context module raises LaSOT AUC from 70.1% (sequence sampling without Mamba) to 71.8%; increasing context token length beyond 1 lowers performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ViT with HiViT (ablation) | OSTrack baseline · LaSOT | AUC: 69.1% -> 70.5% | Helped |
| use sequence sampling instead of single-frame sampling (ablation) | OSTrack baseline · LaSOT | AUC: 69.1% -> 70.1% | No clear effect |
| add Mamba context learning to the sequence-sampling tracker (ablation) | OSTrack with HiViT and sequence sampling · LaSOT | AUC: 70.1% -> 71.8% | Helped |
| set context token length to 2 instead of 1 (ablation) | MambaLCT-256 · LaSOT | AUC: 71.8% -> 71.3% | Hurt |
| set context token length to 3 instead of 1 (ablation) | MambaLCT-256 · LaSOT | AUC: 71.8% -> 70.8% | Hurt |
| set context token length to 4 instead of 1 (ablation) | MambaLCT-256 · LaSOT | AUC: 71.8% -> 69.8% | Hurt |
| remove Mamba layers (ablation) | MambaLCT-256 · LaSOT | AUC: 70.1% -> 71.8% | Hurt |
| insert Mamba in layers (1,2,3) rather than (3,6,9) (ablation) | MambaLCT-256 · LaSOT | AUC: 70.5% -> 71.8% | Hurt |
| insert Mamba in layers (18,19,20) rather than (3,6,9) (ablation) | MambaLCT-256 · LaSOT | AUC: 70.9% -> 71.8% | Hurt |
| use MambaLCT-384 rather than ODTrack | LaSOT · LaSOT | AUC: +0.4% | Helped |
| use MambaLCT-384 rather than AQATrack-384 | LaSOT_ext · LaSOT_ext | AUC: +0.6% | Helped |
| use MambaLCT-384 rather than ARTrack-384 | GOT-10K · GOT-10K | AO: +1.7% | Helped |
Surprising findings
- Increasing the context token length (c_p) from 1 to 4 consistently degraded performance, which the authors attribute to redundant context from Mamba's autoregressive nature.
- Using video sequence sampling alone (without Mamba context) did not improve tracking performance over single-frame sampling.
Limitations
- Training and testing are not unified: during training only short sampled clips (length 2) are used rather than full sequences, due to compute constraints.
- Increasing context token length degrades performance, suggesting redundancy in the context representation.
Experimental setup
Datasets: GOT-10K LaSOT COCO TrackingNet
Benchmarks: LaSOT LaSOT_ext GOT-10K TrackingNet TNL2K UAV123
Metrics: AUC P_norm precision AO SR0.5 SR0.75 speed (fps)
Baselines: ODTrack AQATrack SeqTrack ROMTrack ARTrack OSTrack VideoTrack F-BDMTrack STARK MixFormer
Builds on: OSTrack Vim HiViT ODTrack MAE
Closest papers by meaning
- Robust Tracking via Mamba-based Context-aware Token Learning (AAAI 2025 · similarity 0.90)
- Exploring Enhanced Contextual Information for Video-Level Object Tracking (AAAI 2025 · similarity 0.89)
- Less Is More: Token Context-Aware Learning for Object Tracking (AAAI 2025 · similarity 0.88)
- Mamba-Adaptor: State Space Model Adaptor for Visual Recognition (CVPR 2025 · similarity 0.86)
- ODTrack: Online Dense Temporal Token Learning for Visual Tracking (AAAI 2024 · similarity 0.86)
- Vision Mamba Mender (NeurIPS 2024 · 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.