Research map / Detection and segmentation / Open-set and few-shot detection
Efficient Test-time Adaptive Object Detection via Sensitivity-Guided Pruning
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 an efficient continual test-time adaptive object detection method that prunes domain-sensitive BN channels via sensitivity-guided weighted L1 sparsity plus stochastic channel reactivation, cutting FLOPs ~12% versus the prior SOTA.
Problem
Continual test-time adaptive object detection (CTTA-OD) methods focus on accuracy and largely ignore computational efficiency, which matters for resource-constrained settings; not all source feature channels help under domain shift.
Approach
Quantify each BN channel's sensitivity to domain shift at image and instance levels, apply a sensitivity-weighted L1 sparsity loss on BN scaling factors to prune sensitive channels (building a sub-network per step), and stochastically reactivate pruned channels via Bernoulli sampling to avoid early mis-pruning.
Key results
- On Cityscapes→Cityscapes-C, average 10-round mAP 11.4 vs 9.1 for WHW (+2.3) with total FLOPs 449.6G vs 508.2G (about 11.6% reduction).
- On UAVDT→UAVDT-C, average mAP 14.3 vs 12.0 for WHW (+2.3) with 12.9% lower FLOPs.
- On Cityscapes→ACDC, average mAP 24.2 vs 23.6 for WHW (+0.6) with 10.9% lower FLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply sensitivity-guided channel pruning | CTTA-OD with Faster R-CNN, ResNet18 · Cityscapes → Cityscapes-C | Avg mAP: 9.1 -> 11.4 | Helped |
| Apply sensitivity-guided channel pruning | CTTA-OD with Faster R-CNN, ResNet18 · UAVDT → UAVDT-C | Avg mAP: 12.0 -> 14.3 | Helped |
| Apply sensitivity-guided channel pruning | CTTA-OD with Faster R-CNN, ResNet18 · Cityscapes → ACDC | Avg mAP: 23.6 -> 24.2 | Helped |
| Apply sensitivity-guided channel pruning | CTTA-OD with Faster R-CNN, ResNet18 · Cityscapes → Cityscapes-C | Total FLOPs (G): 508.2 -> 449.6 | Helped |
| Apply sensitivity-guided channel pruning | CTTA-OD with Faster R-CNN, ResNet18 · UAVDT → UAVDT-C | Total FLOPs (G): 508.2 -> 442.4 | Helped |
| Apply sensitivity-guided channel pruning | CTTA-OD with Faster R-CNN, ResNet18 · Cityscapes → ACDC | Total FLOPs (G): 508.2 -> 452.8 | Helped |
| Add unweighted sparsity regularization L_reg (ablation) | L_adp adaptation · Cityscapes → Cityscapes-C | Avg mAP: 10.9 -> 8.8 | Hurt |
| Add image-level sensitivity weighting (ablation) | L_adp with sparsity regularization · Cityscapes → Cityscapes-C | Avg mAP: 8.8 -> 10.5 | Helped |
| Add instance-level sensitivity weighting (ablation) | L_adp with sparsity regularization · Cityscapes → Cityscapes-C | Avg mAP: 8.8 -> 9.9 | Helped |
| Use both image-level and instance-level sensitivity weighting (ablation) | Sensitivity-guided pruning · Cityscapes → Cityscapes-C | Avg mAP: 8.8 -> 11.2 | Helped |
| Add stochastic channel reactivation (ablation) | Sensitivity-guided pruning · Cityscapes → Cityscapes-C | Avg mAP: 11.2 -> 11.4 | Helped |
Surprising findings
- Adapting all channels does not reduce FLOPs, and removing certain source channels can improve cross-domain performance while hurting in-domain performance.
Limitations
- Relies on BN layers, so it does not directly apply to detectors without BN layers (e.g., some transformer detectors).
- Gains on ACDC are small (+0.6 mAP over WHW).
- Pruning hyper-parameters (threshold, ratio, lambda) need tuning and trade off performance against FLOPs.
Experimental setup
Datasets: Cityscapes Cityscapes-C ACDC UAVDT UAVDT-C
Benchmarks: Cityscapes-C ACDC UAVDT-C
Metrics: mAP@50 FLOPs (forward, backward, total)
Baselines: Direct Test TeST DUA STFAR ActMAD WHW
Builds on: Network Slimming (BN scaling L1 sparsity) Faster R-CNN CTTA-OD feature alignment (ActMAD, WHW)
Closest papers by meaning
- Adaptive Object Detection Using Adjacency and Zoom Prediction (CVPR 2016 · similarity 0.84)
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.83)
- DetectoRS: Detecting Objects With Recursive Feature Pyramid and Switchable Atrous Convolution (CVPR 2021 · similarity 0.82)
- As Pseudo-Label Free as Possible: Leveraging Adaptive Feature Generation for Sparsely Annotated Object Detection (AAAI 2025 · similarity 0.81)
- Depth-aware Test-Time Training for Zero-shot Video Object Segmentation (CVPR 2024 · similarity 0.81)
- Learning Endogenous Attention for Incremental Object Detection (CVPR 2025 · similarity 0.80)
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.