SiMHand: Mining Similar Hands for Large-Scale 3D Hand Pose Pre-training

TL;DR

SiMHand tackles scalable pre-training for 3D hand pose estimation by mining similar hands from $2.0M$ in-the-wild images and formulating a similarity-aware contrastive objective. It constructs a large, diverse pre-training set from Ego4D and 100DOH, extracts compact pose embeddings via 2D keypoints and PCA, and uses adaptive weighting in NT-Xent loss to emphasize informative pairs. Across FreiHand, DexYCB, and AssemblyHands, SiMHand achieves state-of-the-art results, validating the benefits of non-self-positive samples and large-scale in-the-wild data for robust 3D hand pose pre-training. The approach demonstrates strong resilience to occlusion and background variation, with practical impact for real-world hand pose understanding in vision and robotics applications.

Abstract

We present a framework for pre-training of 3D hand pose estimation from in-the-wild hand images sharing with similar hand characteristics, dubbed SimHand. Pre-training with large-scale images achieves promising results in various tasks, but prior methods for 3D hand pose pre-training have not fully utilized the potential of diverse hand images accessible from in-the-wild videos. To facilitate scalable pre-training, we first prepare an extensive pool of hand images from in-the-wild videos and design our pre-training method with contrastive learning. Specifically, we collect over 2.0M hand images from recent human-centric videos, such as 100DOH and Ego4D. To extract discriminative information from these images, we focus on the similarity of hands: pairs of non-identical samples with similar hand poses. We then propose a novel contrastive learning method that embeds similar hand pairs closer in the feature space. Our method not only learns from similar samples but also adaptively weights the contrastive learning loss based on inter-sample distance, leading to additional performance gains. Our experiments demonstrate that our method outperforms conventional contrastive learning approaches that produce positive pairs sorely from a single image with data augmentation. We achieve significant improvements over the state-of-the-art method (PeCLR) in various datasets, with gains of 15% on FreiHand, 10% on DexYCB, and 4% on AssemblyHands. Our code is available at https://github.com/ut-vision/SiMHand.

Figures (9)

• Figure 1: The pipeline of pre-training and fine-tuning.(Left) Previous pre-training methods (e.g., PeCLR spurr:iccv21) learn from positive pairs originating from the different augmentations and fine-tune the network on a dataset. (Right) Our method is designed to learn from positive pairs with similar foreground hands, sampled from a pool of hand images in the wild.
• Figure 2: Visualization of similar hand samples in Top-K. Given the query image ($I$), the mined similar samples are shown ("Top-1" corresponds to $I^+$ in Sec. \ref{['sec:method_simhand']}).
• Figure 2: Comparison with different pre-training data sizes. '*' indicates that we use a small amount of training data for fine-tuning to validate the effectiveness of the pre-trained model. Our method demonstrates a leading advantage across all pre-training data scales.
• Figure 3: Overview of our SiMHand. Starting from the left, hand images ($I$, $I^{+}$, $I^{-}$) and their corresponding 2D keypoints are input to the model. After applying random augmentations through transformation $\mathbf {T}$, both the images and 2D keypoints are spatially transformed. The altered 2D keypoints are then used to compute adaptive weights $w_{\text{pos}}$ and $w_{\text{neg}}$, which guide contrastive learning by strengthening or weakening the alignment between positive and negative samples.
• Figure 4: Comparison with different data availability in fine-tuning on FreiHand. Variations in the percentage of labeled data correspond to different subsets of the fine-tuning dataset, following the experimental design in spurr:iccv21.