# Vision Transformer

ViTMatte is an image segmentation system based on a pretrained pure vision transformer (Plain Vision Transformers, ViTs). It optimizes the balance between performance and computational efficiency by utilizing a hybrid attention mechanism and convolutional neck, and introduces a detail capture module to supplement the detailed information required for segmentation. ViTMatte is the first work to harness the potential of ViT in the field of image segmentation through simple adaptation, inheriting the advantages of ViT in terms of pretraining strategy, concise architecture design, and flexible inference strategy. In the Composition-1k and Distinctions-646, the most commonly used image segmentation benchmark tests, ViTMatte achieves state-of-the-art performance and surpasses previous works significantly.

SMPLer-X is a human pose and shape estimation model based on big data and large models. It can unify the capture of body, hand, and facial movements and boasts wide applicability. Through systematic research on datasets from 32 different scenarios, the model optimizes training schemes and dataset selection, leading to a significant enhancement in EHPS capabilities. SMPLer-X employs Vision Transformer for model expansion and utilizes a fine-tuning strategy to transform it into an expert model, further improving performance. The model has demonstrated outstanding performance in multiple benchmark tests, including AGORA (107.2 mm NMVE), UBody (57.4 mm PVE), EgoBody (63.6 mm PVE), and EHF (62.3 mm PVE without finetuning). SMPLer-X's strength lies in its ability to handle diversified data sources, exhibiting excellent generalization capability and transferability.