Source-linked AI summary
Progressive Domain Adaptation for Object Detection
Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, Ming-Hsuan Yang
TL;DR
Object detection needs costly bounding-box annotations and can generalize poorly across different data distributions. The paper translates source images into an intermediate domain, progressively aligns features through two easier subtasks, and weights synthetic samples by quality; experiments report favorable target-domain performance against the state of the art across varied scenarios.
Problem
Object detection models require large annotated datasets and often generalize poorly when training and testing domains differ, while target-domain annotation is costly.
Method
The method creates a synthetic intermediate domain by transforming source images, then progressively aligns source-to-intermediate and intermediate-to-target features with a quality-weighted task loss.
Results
The method performs favorably against the state-of-the-art method across weather, camera, and large-scale-dataset adaptation scenarios.
Takeaways & Limitations
Progressive adaptation reduces domain discrepancy by decomposing adaptation into smaller-gap subtasks and accounting for uneven synthetic-image quality.
Takeaways & Limitations
Synthetic-image quality varies widely, with some translations losing object details or introducing artifacts and lying far from both source and target domains.
Abstract
from arXiv · showhide
Recent deep learning methods for object detection rely on a large amount of bounding box annotations. Collecting these annotations is laborious and costly, yet supervised models do not generalize well when testing on images from a different distribution. Domain adaptation provides a solution by adapting existing labels to the target testing data. However, a large gap between domains could make adaptation a challenging task, which leads to unstable training processes and sub-optimal results. In this paper, we propose to bridge the domain gap with an intermediate domain and progressively solve easier adaptation subtasks. This intermediate domain is constructed by translating the source images to mimic the ones in the target domain. To tackle the domain-shift problem, we adopt adversarial learning to align distributions at the feature level. In addition, a weighted task loss is applied to deal with unbalanced image quality in the intermediate domain. Experimental results show that our method performs favorably against the state-of-the-art method in terms of the performance on the target domain.
1. Introduction
Object detection models require costly annotations and often generalize poorly across domains. The paper proposes an intermediate synthetic domain and weighted adaptation to decompose a large domain shift into easier alignment problems.
- Object detection models depend on large annotated datasets, yet domain differences in scenes, weather, lighting, and cameras impair generalization.Additional target-domain annotations can improve performance but are time-consuming and labor-intensive to collect.
- The method transforms source images to resemble target images, creating an intermediate domain between the source and target distributions.This intermediate domain supports progressive alignment rather than direct mapping across a large distribution gap.
- Progressive adaptation decomposes source-to-target alignment into two subtasks with smaller gaps: source-to-intermediate and intermediate-to-target.The intermediate domain is intended to make each adaptation step easier than direct source-to-target alignment.
- Synthetic images receive unequal weights according to their quality relative to the target domain, reducing the influence of low-quality translations.The weight is based on distance to the target distribution and is incorporated into the task loss.
- Across multiple datasets and domain discrepancies, the method performs favorably against the state-of-the-art algorithm on target-domain accuracy.Evaluated scenarios include weather changes, camera differences, and adaptation to a large-scale dataset.
2. Related Work
Prior work uses supervised, synthetic-data, adversarial, and image-level strategies for domain adaptation, but object detection remains challenging under substantial domain shifts. This paper combines an intermediate synthetic domain with two-stage alignment and sample weighting.
- Deep object detectors achieve strong performance but rely on substantial labeled training data and may overfit the training domain.These properties make generalization to other domains difficult and impose high annotation costs.
- Domain adaptation methods seek to align source and target distributions using unlabeled or weakly labeled target images, often through adversarial feature alignment.Prior work has focused more heavily on image classification than on complex tasks such as object detection.
- Existing object-detection approaches include limited-box fine-tuning, synthetic-data training with target pseudo-labels, and image- or instance-level adversarial alignment.These approaches address weakly supervised or unsupervised adaptation settings in different ways.
- The framework uses source-to-intermediate and intermediate-to-target subtasks to progressively close the domain gap.The intermediate synthetic domain is positioned between the source and target distributions.
- The paper argues that image translation without measuring each generated sample’s target distance can produce less effective adaptation.Its response is an intermediate domain, two-stage alignment, and sample weights reflecting synthetic-image quality.
3. Progressive Domain Adaptation
The method bridges source and target domains with a synthetic intermediate domain, progressively aligning source-to-intermediate and intermediate-to-target distributions. It uses adversarial feature alignment and weights synthetic samples according to their proximity to the target distribution.
- Progressive Adaptation: Progressive adaptation decomposes direct source-to-target alignment into two subtasks, S →F and F →T, bridged by a synthetic domain.The synthetic domain is constructed from source images translated toward the target appearance.
- Feature-Space Adaptation: The model aligns feature distributions with a detection network and a discriminator trained through adversarial learning.A gradient reversal layer makes the encoder produce features that fool the discriminator while the discriminator distinguishes domains.
- Progressive Adaptation: The intermediate domain reduces the difficulty of adapting across a large domain gap by progressively aligning distributions rather than solving the full gap directly.The approach focuses on individual discrepancies during each adaptation stage.
- Progressive Adaptation: t-SNE visualization places synthetic features between the KITTI source and Cityscapes target distributions, supporting the intermediate-domain interpretation.The visualization compares 500 Cityscapes validation images with 500 KITTI training images.
- Weighted Supervision: Synthetic-image quality varies, with artifacts and lost object details producing outliers that can cause false feature alignment.The method assigns lower importance to synthetic images farther from the target distribution and applies the weights to the second-stage detection loss.
4. Experimental Results
Experiments evaluate progressive adaptation across cross-camera, weather, and large-scale dataset shifts, using baselines, oracle models, and ablations. The method improves target-domain detection, with gains from synthetic data, progressive training, and importance weighting.
- Experimental setup: Experiments cover cross-camera, weather, and large-scale dataset adaptation scenarios, comparing source-only baselines, target-supervised oracle models, and the proposed method.The datasets include KITTI, Cityscapes, Foggy Cityscapes, and BDD100k.
- Cross Camera Adaptation: Table 1 reports average precision for the shared car class in cross-camera adaptation using KITTI and Cityscapes.
- Weather Adaptation: 10% mAP improvement over the baseline occurs when synthetic images are introduced during progressive weather adaptation from Cityscapes to Foggy Cityscapes.The synthetic target domain is closely distributed to the foggy target and inherits informative labels from the source images.
- Adaptation to Large-scale Dataset: 3.1% mAP gain over the baseline is achieved when progressively adapting from Cityscapes to the diverse BDD100k daytime target domain.Direct source-to-target alignment improves performance by 0.4%, while adding synthetic data improves it by 2.5%.
5. Conclusions
The method bridges source and target domains through an intermediate domain, progressively adapting between distributions while balancing image-quality differences. Experiments report favorable performance against a state-of-the-art method across varied domain-shift scenarios.
- The method transforms source images toward the target appearance and progressively adapts through an intermediate domain.It first adapts from source to intermediate data, then from the intermediate domain to the target domain.
- A weighted loss balances different image qualities in the intermediate domain during the second adaptation stage.
- The method performs favorably against the state-of-the-art method across cross-camera, weather, and large-scale-dataset adaptation scenarios.