IC-647DINOv2's feature-map artifacts cause it to be incompatible with the LOSt unsupervised object discovery method, scoring far below DINO

Timothée Darcet, Maxime Oquab, Julien Mairal, Piotr Bojanowski

SourceVision Transformers Need Registers

The LOSt algorithm, which detects objects by gathering information in attention maps, works well on DINO features (61.9 corloc on VOC 2007) but produces degraded performance on DINOv2 features. The paper reports DINOv2 achieving only 35.3/40.2/26.9 corloc on VOC 2007/2012 and COCO 20k, and DeiT-III even lower at 11.7/13.1/10.7. The authors attribute this to the high-norm outlier tokens corrupting the local feature maps that LOSt relies on. The qualitative analysis in Appendix C shows that the LOSt seed expansion score is disrupted by the artifacts in DINOv2 and DeiT-III, while for OpenCLIP the value projection appears to filter out the outliers, explaining its relatively better LOSt performance.

Evidence
observational
Key metric
LOSt corloc: DINOv2 35.3/40.2/26.9 (VOC 2007/2012, COCO 20k); DeiT-III 11.7/13.1/10.7; OpenCLIP 38.8/44.3/31.0; DINO 61.9 (VOC 2007, from Simeoni et al.)
Caveat
The LOSt numbers in Table 3 are from the authors' reproductions of the training recipes, not the exact released checkpoints. The DINO baseline of 61.9 is taken from Simeoni et al. 2021.
Model
DINOv2, DINO, DeiT-III, OpenCLIP
Concepts
Failure mode
Datasets
MS COCO / COCO / COCO 2014 / COCO 2017 / COCO 20k / COCO-it / COCO-wl [eval]
Methods
LOSt [eval]
Related work
LOSt [builds-on]
Related findings
IC-646
Extraction
automatic-extraction