Research map / Architectures and efficiency / Pruning and efficient tuning
"Tracking Meets LoRA: Faster Training, Larger Model, Stronger Performance"
Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, Haibin Ling
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 token type embedding with shared (sliced) positional embedding for template and search (ablation) | LoRAT-L-224 · LaSOT-ext | SUC: 51.9 -> 52.7 | Improved |
| replace MLP-only head with convolutional head under LoRA fine-tuning (ablation) | LoRAT-L-224 · LaSOT | SUC: 74.2 -> 4.7 | Hurt |
| convolutional head vs MLP head under full fine-tuning (ablation) | LoRAT-L-224 · LaSOT | SUC: 73.0 -> 71.7 | Hurt |
| use LoRA instead of full fine-tuning (ablation) | LoRAT-L-224 · LaSOT | SUC: 73.0 -> 74.2 | Improved |
| add foreground indication embedding at 378 resolution (ablation) | LoRAT-L-378 · LaSOT-ext | SUC: 55.2 -> 56.6 | Improved |
| unfreeze positional embedding (ablation) | LoRAT-L-224 · LaSOT | SUC: 73.5 vs 73.9 | No clear effect |
| use DINOv2 pre-training instead of MAE, CLIP or EVA-02 (ablation) | LoRAT-B-224 (full fine-tuning) · LaSOT | SUC: 70.3 (MAE) -> 70.9 | Improved |
| use LoRA on all linear layers (attention and MLP) rather than only q,v (ablation) | LoRAT-L-224 · LaSOT | SUC: 71.9 -> 74.2 | Improved |
| replace LoRA with Adapter, VPT or IA3 (ablation) | LoRAT-L-224 · LaSOT | SUC: 74.2 -> 55.0 (Adapter) | Hurt |
| use truncated normal init for LoRA instead of Kaiming uniform or Gaussian (ablation) | LoRAT-L-224 · LaSOT | SUC: 73.7 -> 74.2 | Improved |
| replace full self-attention with S-MAM (ablation) | LoRAT-L-224 · LaSOT | SUC: 74.2 -> 73.5 | Hurt |
| use LoRA instead of full fine-tuning for training efficiency (ablation) | LoRAT-g-378 · Training memory (8 V100) | GPU memory / feasibility: Out of Memory -> 25.8 GB | Improved |
Closest papers by meaning
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.87)
- ExPLoRA: Parameter-Efficient Extended Pre-Training to Adapt Vision Transformers under Domain Shifts (ICML 2025 · similarity 0.87)
- Track-On: Transformer-based Online Point Tracking with Memory (ICLR 2025 · similarity 0.86)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.86)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.86)
- Is Multiple Object Tracking a Matter of Specialization? (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.
