Unsupervised self-training correction learning for 2D image-based 3D model retrieval

Yaqian Zhou, Yu Liu, Jun Xiao, Min Liu, Xuanya Li, An-An Liu

Abstract: Highlights•We propose an unsupervised self-training correction learning (USTCL) framework for 2D image-based 3D model retrieval (IBMR).•We utilize a noise-corrected self-training learning module to denoise pseudo labels to make the predicted categories more easily discriminated, thereby improving the prediction discriminability.•We adopt a target-guided pseudo label refining module to refine pseudo labels to prevent minority categories from being pushed into majority categories, thereby improving prediction diversity.•Comprehensive experiments on the popular IBMR benchmarks, MI3DOR and MI3DOR-2, validate that USTCL outperforms the existing competing methods.