3D human pose estimation remains a challenge due to depth ambiguity, occlusion, and temporal inconsistency within a video sequence. Existing coordinate regression methods have limitations in predicting unrealistic poses in complex situations, either o...
3D human pose estimation remains a challenge due to depth ambiguity, occlusion, and temporal inconsistency within a video sequence. Existing coordinate regression methods have limitations in predicting unrealistic poses in complex situations, either overlooking inter-joint dependencies or being sensitive to noise. To address this, this paper proposes a new framework that extends the concept of
previous work to the space-time region and compresses pose information into discrete space and time tokens. The proposed model consists of two steps. The first step is to learn a
tokenizer that receives a 3D sequence and discretizes the spatial partial structure and temporal flow through each codebook. The second step is to perform the classification problem of predicting the previously learned token index from 2D joint input, and for this purpose, we designed a classifier based on ST-GCN. Experiments on the Human3.6M dataset show that the proposed model performed on par with the existing
state-of-the-art methodology, and demonstrated robust recovery capability, especially in environments with severe masking. In addition, ablation studies confirmed that long sequences and the appropriate number of token classes contribute to fine motion capture. In addition, after encoding joint information,
the proposed method showed better performance to go through the time information processing process. This study presents a new direction for securing stability and temporal consistency of estimation by reinterpreting posture estimation as a space-time token classification problem rather than a continuous regression.