Domain-Adaptive 2D Human Pose Estimation via Dual Teachers in Extremely Low-Light Conditions

Existing 2D human pose estimation research predominantly concentrates onwell-lit scenarios, with limited exploration of poor lighting conditions, whichare a prevalent aspect of daily life. Recent studies on low-light poseestimation require the use of paired well-lit and low-light images with groundtruths for training, which are impractical due to the inherent challengesassociated with annotation on low-light images. To this end, we introduce anovel approach that eliminates the need for low-light ground truths. Ourprimary novelty lies in leveraging two complementary-teacher networks togenerate more reliable pseudo labels, enabling our model achieves competitiveperformance on extremely low-light images without the need for training withlow-light ground truths. Our framework consists of two stages. In the firststage, our model is trained on well-lit data with low-light augmentations. Inthe second stage, we propose a dual-teacher framework to utilize the unlabeledlow-light data, where a center-based main teacher produces the pseudo labelsfor relatively visible cases, while a keypoints-based complementary teacherfocuses on producing the pseudo labels for the missed persons of the mainteacher. With the pseudo labels from both teachers, we propose aperson-specific low-light augmentation to challenge a student model in trainingto outperform the teachers. Experimental results on real low-light dataset(ExLPose-OCN) show, our method achieves 6.8% (2.4 AP) improvement over thestate-of-the-art (SOTA) method, despite no low-light ground-truth data is usedin our approach, in contrast to the SOTA method. Our code will be availableat:https://github.com/ayh015-dev/DA-LLPose.