Research map / Architectures and efficiency / Attention and transformers

Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking

CVPR 2025 · Architectures and efficiency › Attention and transformers · approach: Hybrid and interpretable ViT designs · 56 citations (OpenAlex)

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 SGLATrack, a lightweight ViT UAV tracker that disables redundant later ViT layers after a saturated layer and keeps one layer chosen by a selection module trained with a layer-wise similarity loss.

Problem

Vision transformer trackers are too heavy for real-time UAV tracking, and layer redundancy in lightweight ViT trackers is unexplored.

Approach

Observes that search features saturate in deeper layers of tiny ViTs; after a saturated layer (l*=6), most subsequent layers are disabled and a single layer is retained via an MLP selection module trained to maximize cosine similarity with the saturated features.

Key results

• SGLATrack-DeiT* reaches 66.9% AUC on UAV123 at about 224.7 FPS on GPU (GTX 2080Ti) and 74.8 FPS on CPU.
• On UAVTrack112, SGLATrack-DeiT* outperforms second-best AVTrack by 2.1% AUC and 2.5% precision.
• Layer adaptation cuts parameters by about 27% and FLOPs by about 30% with at most 0.4% AUC loss (SGLATrack-ViT on UAVTrack112); runs at 33 FPS on Jetson TX2.

Measured effects reported in the paper

Surprising findings

• Pruning deeper layers directly causes a large precision drop; the selection module that retains one similar layer is needed to recover accuracy.
• Maximizing similarity to the saturated layer outperforms minimizing it (65.7 vs 66.9 AUC on UAV123).

Limitations

• Precision on some datasets (UAVDT, UAVTrack112 L) lags trackers with lower AUC, attributed to small sequence counts.
• Saturated layer l* is set via prior knowledge (fixed to 6); the adaptive threshold alternative is slower.
• Evaluated only on tiny ViT backbones and UAV-focused benchmarks.

Experimental setup

Datasets: LaSOT COCO TrackingNet GOT-10k

Benchmarks: DTB70 UAVDT UAV123 UAV123@10fps UAVTrack112 UAVTrack112 L

Metrics: AUC precision FPS

Baselines: AVTrack-DeiT SMAT HiT-Small MVT TCTrack++ TCTrack UDAT HCAT LightTrack HiFT AutoTrack ARCF

Builds on: OSTrack AVTrack DeiT EVA focal loss

Closest papers by meaning

• Learning Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking (ICML 2024 · similarity 0.92)
• Learning Occlusion-Robust Vision Transformers for Real-Time UAV Tracking (CVPR 2025 · similarity 0.88)
• LVPTrack: High Performance Domain Adaptive UAV Tracking with Label Aligned Visual Prompt Tuning (AAAI 2025 · similarity 0.85)
• Target Focused Shallow Transformer Framework for Efficient Visual Tracking (AAAI 2024 · similarity 0.84)
• Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.84)
• A General and Efficient Training for Transformer via Token Expansion (CVPR 2024 · similarity 0.84)