Vision Pretraining for Dense Spatial Perception

Open paper
Benchmark Model Rank Results
linear-probe-semantic-segmentation-on-ade20kLingBot-Vision ViT-g + linear probe#2mIoU: 53.5
linear-probe-semantic-segmentation-on-ade20kLingBot-Vision ViT-L + linear probe#3mIoU: 52.75
linear-probe-semantic-segmentation-on-ade20kLingBot-Vision ViT-B + linear probe#4mIoU: 51.44
linear-probe-semantic-segmentation-on-ade20kLingBot-Vision ViT-S + linear probe#9mIoU: 47.01
frozen-feature-label-propagation-on-davis-2017-valLingBot-Vision ViT-g + label propagation#3J&F: 70.0
image-classification-on-imagenetLingBot-Vision ViT-L (linear probe)#163Top 1 Accuracy: 86.38%
image-classification-on-imagenetLingBot-Vision ViT-g (linear probe)#165Top 1 Accuracy: 86.32%
image-classification-on-imagenetLingBot-Vision ViT-B (linear probe)#280Top 1 Accuracy: 85.05%
image-classification-on-imagenetLingBot-Vision ViT-S (linear probe)#593Top 1 Accuracy: 82.22%

Rank counts only results with a code link.