Yin Yin

Greg Heinrich, Mike Ranzinger, Hongxu et al.

Agglomerative models have recently emerged as a powerful approach to training vision foundation models, leveraging multi-teacher distillation from existing models such as CLIP, DINO, and SAM. This strategy enables the efficient creation of robust models, combining the strengths of individual teachers while significantly reducing computational and resource demands. In this paper, we thoroughly analyze state-of-the-art agglomerative models, identifying critical challenges including resolution mode shifts, teacher imbalance, idiosyncratic teacher artifacts, and an excessive number of output tokens. To address these issues, we propose several novel solutions: multi-resolution training, mosaic augmentation, and improved balancing of teacher loss functions. Specifically, in the context of Vision Language Models, we introduce a token compression technique to maintain high-resolution information within a fixed token count. We release our top-performing variants at multiple scales (-B, -L, -H, and -g), along with inference code and pretrained weights

Yin Yin, Sergei V. Fotin, Hrishikesh Haldankar et al.

A novel hierarchical model is introduced to solve a general problem of detecting groups of similar objects. Under this model, detection of groups is performed in hierarchically organized layers while each layer represents a scope for target objects. The processing of these layers involves sequential extraction of appearance features for an individual object, consistency measurement features for nearby objects, and finally the distribution features for all objects within the group. Using the concept of scope-based normalization, the extracted features not only enhance local contrast of an individual object, but also provide consistent characterization for all related objects. As an example, a microcalcification group detection system for 2D mammography was developed, and then the learned model was transferred to 3D digital breast tomosynthesis without any retraining or fine-tuning. The detection system demonstrated state-of-the-art performance and detected 96% of cancerous lesions at the rate of 1.2 false positives per volume as measured on an independent tomosynthesis test set.