IC-1619DINOv2's patch-level features outperform CLIP and MAE for cross-image semantic feature matching

Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, Chunhua Shen

SourceMatcher: Segment Anything with One Shot Using All-Purpose Feature Matching

Using each model as the image encoder in the Matcher framework, the paper measures how well their patch-level features support cross-image semantic matching. DINOv2 (ViT-L/14) achieves 52.7% mean mIoU on COCO-20i, 87.0% on FSS-1000, and 79.5 J&F on DAVIS 2017, substantially outperforming CLIP (32.2/77.4/73.9) and MAE (18.8/71.9/69.5). The authors attribute CLIP's weakness to text-image contrastive pre-training limiting pixel-level information, and MAE's weakness to patch-level features confusing surrounding patch information. DINOv2's image-level and patch-level discriminative self-supervised learning is credited with producing all-purpose visual features suitable for precise patch matching.

Evidence
correlational
Key metric
DINOv2: 52.7 / 87.0 / 79.5; CLIP: 32.2 / 77.4 / 73.9; MAE: 18.8 / 71.9 / 69.5 (mean mIoU COCO-20i, mIoU FSS-1000, J&F DAVIS 2017)
Caveat
Performance is measured within the Matcher framework's specific matching protocol (cosine similarity + Hungarian matching); the relative ranking may differ under other matching or evaluation setups.
Model
DINOv2, CLIP / CLIP-ViT (LC), MAE
Concepts
Distance preservation
Datasets
COCO-20i [eval], FSS-1000 [eval], DAVIS 2017 val [eval]
Related findings
IC-1618
Extraction
automatic-extraction