Research map / Detection and segmentation / Open-set and few-shot detection
ISP-Teacher:Image Signal Process with Disentanglement Regularization for Unsupervised Domain Adaptive Dark Object Detection
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 ISP-Teacher, a Teacher-Student domain adaptive dark object detector that replaces non-learnable night augmentation with a learnable ISP-based day-to-night transformation and adds a disentanglement regularization on gradient cosine similarity.
Problem
Unsupervised domain adaptive object detection from daytime to nighttime images, where Teacher-Student methods degrade under day-to-night shift because non-learnable augmentations ignore the camera ISP imaging process.
Approach
Model the camera ISP pipeline (invert processing, shot/read noise, ISP restoration) as a day-to-night transformation whose parameters (light intensity, gamma, white balance gains) are learned self-supervisedly by an encoder-decoder, and regularize the shared encoder with a gradient cosine-similarity loss that pushes detection and ISP-degradation gradients toward orthogonality.
Key results
- On BDD100k, ISP-Teacher reaches 48.8 AP versus 46.4 for 2PCNet (+2.4) and 41.1 for the source-only lower bound.
- On SHIFT, ISP-Teacher reaches 52.4 AP versus 49.1 for 2PCNet (+3.3).
- Ablation on BDD100k: ISP degradation self-supervision raises AP from 42.2 (plain augmentation) to 48.5; adding DR gives 48.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ISP-Teacher instead of 2PCNet | Dark object detection · BDD100k night val | AP: 46.4 -> 48.8 | Helped |
| Use ISP-Teacher instead of 2PCNet | Dark object detection · SHIFT night val | AP: 49.1 -> 52.4 | Helped |
| Replace non-learnable daytime augmentation with self-supervised ISP degradation (ablation) | ISP-Teacher · BDD100k night val | AP: 42.2 -> 48.5 | Helped |
| Add NightAug (ablation) | Faster-RCNN domain adaptation · BDD100k night val | AP: 42.2 -> 45.8 | Helped |
| Add disentanglement regularization (ablation) | ISP-Teacher with self-supervised ISP degradation · BDD100k night val | AP: 48.5 -> 48.8 | Helped |
| Increase self-supervised loss weight β from 1 to 2 (ablation) | ISP-Teacher · BDD100k night val | AP: 48.8 -> 48.4 | Hurt |
| Increase self-supervised loss weight β from 1 to 5 (ablation) | ISP-Teacher · BDD100k night val | AP: 48.8 -> 44.9 | Hurt |
Surprising findings
- Teacher-Student methods AT and TDD perform below the source-only Faster R-CNN baseline on BDD100k day-to-night transfer.
- Increasing the self-supervised loss weight beta to 5 drops AP to 44.9, below the non-learnable NightAug baseline (45.8).
Limitations
- Small objects such as traffic lights still perform poorly
- Performance is sensitive to the self-supervised loss weight beta
- Evaluated only on two driving datasets with Faster R-CNN ResNet50
Experimental setup
Datasets: BDD100k SHIFT
Benchmarks: BDD100k SHIFT
Metrics: AP AP50 APS APM APL
Baselines: Source (Faster R-CNN lower bound) Oracle (upper bound) UMT TDD AT 2PCNet
Builds on: Mean Teacher / Teacher-Student UDA (Sohn et al. 2020) Faster R-CNN 2PCNet MAET PCGrad
Closest papers by meaning
- AdaptiveISP: Learning an Adaptive Image Signal Processor for Object Detection (NeurIPS 2024 · similarity 0.86)
- Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation (CVPR 2024 · similarity 0.85)
- D3T: Distinctive Dual-Domain Teacher Zigzagging Across RGB-Thermal Gap for Domain-Adaptive Object Detection (CVPR 2024 · similarity 0.84)
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 2024 · similarity 0.84)
- A-Teacher: Asymmetric Network for 3D Semi-Supervised Object Detection (CVPR 2024 · similarity 0.83)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.83)
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.