Research map / Detection and segmentation / CNN features
What How and When Should Object Detectors Update in Continually Changing Test Domains?
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 adaptation method for object detectors that updates only lightweight adaptor modules, aligns image- and class-wise object features with EMA statistics, and skips adaptation when the test domain is already well adapted.
Problem
Test-time adaptation for object detection under continually changing test domains is underexplored; existing methods update full parameters or BN layers, which is inefficient, forgets task knowledge, and does not fit Transformers or object-level alignment.
Approach
Insert parallel lightweight down/up-projection adaptors (MLP for Transformer blocks, 1x1 conv for ResNet blocks) and update only them; align image-level and class-wise region-level feature statistics with training statistics via EMA; and decide online when to skip backward passes using two distribution-gap criteria.
Key results
- On COCO to COCO-C with Swin-T, Ours improves average mAP by 4.9 points over Direct-Test (22.6 vs 17.7).
- Ours-Skip with ResNet50 cuts backward passes by up to 90.5% and raises FPS from 25.8 to 21.2 while keeping mAP near Ours (21.9 vs 22.2 average).
- On SHIFT-Discrete, Ours improves mAP by 7-9 points on foggy and dawn domains relative to Direct-Test, with the adaptor using 0.5%-0.9% of backbone parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use lightweight adaptors and image- plus object-level feature alignment instead of direct inference | Faster R-CNN with Swin-T · COCO → COCO-C | mAP: 17.7 -> 22.6 | Helped |
| Use lightweight adaptors and image- plus object-level feature alignment instead of direct inference | Faster R-CNN with ResNet50 · COCO → COCO-C | mAP: 17.3 -> 22.2 | Helped |
| Use lightweight adaptors and feature alignment instead of direct inference | Faster R-CNN with Swin-T · SHIFT-Discrete | mAP: 38.5 -> 40.4 | Helped |
| Use lightweight adaptors and feature alignment instead of direct inference | Faster R-CNN with ResNet50 · SHIFT-Discrete | mAP: 35.9 -> 38.7 | Helped |
| Use lightweight adaptors and feature alignment instead of direct inference | Faster R-CNN with Swin-T · SHIFT-Continuous clear↔fog | mAP: 18.1 -> 23.9 | Helped |
| Use lightweight adaptors and feature alignment instead of direct inference | Faster R-CNN with ResNet50 · SHIFT-Continuous clear↔fog | mAP: 12.1 -> 20.9 | Helped |
| Add object-level class-wise feature alignment to image-level alignment (ablation) | Ours loss, Swin-T · COCO → COCO-C | mAP: 16.7 -> 22.6 | Helped |
| Add object-level class-wise feature alignment to image-level alignment (ablation) | Ours loss, ResNet50 · COCO → COCO-C | mAP: 18.1 -> 22.2 | Helped |
| Add class-frequency weighting to object-level alignment (ablation) | Ours loss, Swin-T · COCO → COCO-C | mAP: 17.8 -> 22.6 | Helped |
| Add class-frequency weighting to object-level alignment (ablation) | Ours loss, ResNet50 · COCO → COCO-C | mAP: 18.9 -> 22.2 | Helped |
| Add class-frequency weighting to object-level alignment (ablation) | Ours loss, Swin-T · SHIFT-Continuous clear↔rain | mAP: 25.1 -> 23.2 | Hurt |
| Use adaptor updates rather than full-parameter updates (ablation) | Faster R-CNN with Swin-T · SHIFT-Discrete | mAP: 38.4 -> 40.4 | Helped |
Surprising findings
- Full-parameter updates (ActMAD, Mean-Teacher) cause catastrophic forgetting, dropping performance on the original clear/COCO domain below Direct-Test.
- Updating only LayerNorm or BatchNorm affine parameters gives little gain over full updates on SHIFT for object detection, unlike in classification.
Limitations
- Relies on pre-computed training feature statistics from 2,000 training samples, since training data is inaccessible at test time.
- Skipping thresholds (tau1, tau2) are hand-set and shown to trade off mAP and FPS; evaluated only on COCO-C and SHIFT driving data.
Experimental setup
Datasets: COCO COCO-C SHIFT
Benchmarks: COCO to COCO-C SHIFT-Discrete SHIFT-Continuous
Metrics: mAP FPS number of forward passes number of backward passes trainable parameters memory usage
Baselines: Direct-Test ActMAD Mean-Teacher NORM DUA Evenly-Skip
Builds on: Faster R-CNN FPN Swin Transformer ActMAD TeST NORM DUA Adapter tuning
Closest papers by meaning
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.89)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation (AAAI 2025 · similarity 0.86)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.86)
- Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework (AAAI 2024 · similarity 0.85)
- Test-Time Model Adaptation with Only Forward Passes (ICML 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.