Universal Image Embedding on UnED 1.0 (test)
95.7P@5 (Food2k)IN Specialist+Oracle
Evaluation Results
| Method | Links | ||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| IN Specialist+OraclePre-training=ImageNet21k, Evaluation Protocol=Specialist+Oracle, Backbone=ViT-Base/162024.06 | 95.7 | 50.8 | 81.2 | 68 | 86.2 | 67.3 | 73 | 27.6 | 32.9 | 35.1 | 46.6 | 69.7 | 84.4 | 69.7 | 84.4 | 84.4 | 69.7 | 84.4 | |
| CLIP Specialist+OraclePre-training=CLIP, Evaluation Protocol=Specialist+Oracle, Backbone=ViT-Base/162024.06 | 95.7 | 50.8 | 81.2 | 68 | 86.2 | 67.3 | 73 | 27.6 | 32.9 | 35.1 | 46.6 | 69.7 | 84.4 | 69.7 | 84.4 | 84.4 | 73.6 | 87.1 | |
| UDONPre-training=CLIP, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 50.3 | 62.4 | 80 | 85.8 | 67 | 82.1 | 71.8 | 89.7 | 66.7 | 72.7 | 15.8 | 19.6 | 30.9 | 43.4 | 72.7 | 85.9 | 56.9 | 67.7 | |
| UJCRRPre-training=CLIP, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 50.1 | 62 | 80 | 85.4 | 68.6 | 82.7 | 77 | 91.1 | 63.7 | 69.5 | 4.6 | 5.8 | 25.5 | 36 | 70.1 | 84.1 | 55 | 64.6 | |
| UJCRRPre-training=ImageNet21k, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 49.9 | 62.8 | 61.9 | 71.8 | 60.9 | 78.1 | 66.3 | 85.9 | 70.1 | 75.2 | 20.4 | 24.9 | 31.2 | 43.1 | 73.6 | 87.1 | 54.3 | 66.1 | |
| USCSSPre-training=CLIP, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 49.8 | 62 | 76.4 | 83.4 | 65.8 | 81.3 | 71 | 88.5 | 65.3 | 71.4 | 9.9 | 12.7 | 31.5 | 42.8 | 70.1 | 84.8 | 55 | 65.9 | |
| UDONPre-training=ImageNet21k, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 49.6 | 62.2 | 61.3 | 71.2 | 64.2 | 80.2 | 69.8 | 88.5 | 70.4 | 75.3 | 12 | 15.9 | 28.6 | 40.9 | 75.6 | 88 | 53.9 | 65.3 | |
| USCRRPre-training=CLIP, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 49.5 | 61.4 | 79 | 84.9 | 65.6 | 81.1 | 73.1 | 89.4 | 64.4 | 70.5 | 8.6 | 10.8 | 25.3 | 36.5 | 71.1 | 85.1 | 54.6 | 64.9 | |
| USCSSPre-training=ImageNet21k, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 49 | 61.7 | 53.4 | 64.3 | 62 | 78.8 | 67.6 | 87.2 | 68.3 | 73.5 | 8.4 | 10.7 | 28 | 40.6 | 73.5 | 87.1 | 51.3 | 63 | |
| USCRRPre-training=ImageNet21k, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 48.3 | 60.9 | 58.9 | 69.7 | 61.9 | 78.7 | 70.4 | 88.3 | 69.1 | 74.2 | 7.3 | 9.7 | 21.3 | 31.4 | 74.1 | 87.1 | 51.4 | 62.5 | |
| MLP baseline w/ Dyn SamplerPre-training=ImageNet21k, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 47.5 | 60.1 | 51.3 | 63 | 61.8 | 78.5 | 64.1 | 85 | 69.7 | 74.8 | 14.3 | 17.8 | 25.8 | 36.4 | 73.3 | 86.8 | 51 | 62.8 | |
| USC w/ Dyn SamplerPre-training=ImageNet21k, Descriptor Dimension=64-D, Backbone=ViT-Base/162024.06 | 46.2 | 59.1 | 56.3 | 67.4 | 61.1 | 78.4 | 65.9 | 86.1 | 69.6 | 74.8 | 12 | 15.6 | 25.9 | 37.3 | 73.2 | 86.9 | 51.3 | 63.2 | |
| CLIPPre-training=CLIP, Descriptor Dimension=768-D, Evaluation Protocol=Off-the-shelf, Backbone=ViT-Base/162024.06 | 41.4 | 54.1 | 43.7 | 52.3 | 82.2 | 44.2 | 65.4 | 37.2 | 56 | 52.4 | 61.9 | 21.4 | 28.5 | 20.4 | 31 | 38.6 | 59.9 | 39.8 | |
| SigLIPPre-training=SigLIP, Descriptor Dimension=768-D, Evaluation Protocol=Off-the-shelf, Backbone=ViT-Base/162024.06 | 39.5 | 52.8 | 93.9 | 69.7 | 53.5 | 73.1 | 59.6 | 67.5 | 31.6 | 41.2 | 20.6 | 32 | 42.7 | 64.3 | 49 | 62 | 49 | 62 | |
| DINOv2Pre-training=DINOv2, Descriptor Dimension=768-D, Evaluation Protocol=Off-the-shelf, Backbone=ViT-Base/162024.06 | 35.6 | 51.4 | 67.1 | 56 | 17.4 | 33.4 | 71.2 | 77.6 | 38.3 | 48.1 | 35.4 | 51.7 | 46.6 | 67.8 | 43.9 | 58.2 | 43.9 | 53.5 | |
| INPre-training=ImageNet21k, Descriptor Dimension=768-D, Evaluation Protocol=Off-the-shelf, Backbone=ViT-Base/162024.06 | 31.1 | 44.1 | 35.5 | 53.9 | 67.1 | 74.2 | 36.6 | 53.2 | 40.2 | 68.2 | 21.1 | 30.8 | 14.8 | 25.2 | 52.9 | 74.3 | 38.4 | 52.8 | |
| IN + MIMPre-training=ImageNet21k + Masked Image Modeling, Descriptor Dimension=768-D, Evaluation Protocol=Off-the-shelf, Backbone=ViT-Base/162024.06 | 30.1 | 44.4 | 33 | 51.2 | 53.9 | 62.2 | 32.5 | 50.2 | 38.3 | 68.2 | 20.2 | 29.4 | 42.9 | 74.7 | — | 43.7 | 17.8 | 60.7 |