Sept. 26, 2023, 1:33 p.m. — Public
Win-Win: Training High-Resolution Vision Transformers from Two Windows. Leroy et al. ICLR 2024.
1px total |
1px low-detail |
1px high-detail |
1px matched |
1px unmatched |
1px rigid |
1px non-rigid |
1px not sky |
1px sky |
1px s0-10 |
1px s10-40 |
1px s40+ |
---|---|---|---|---|---|---|---|---|---|---|---|
5.371 | 5.003 | 63.211 | 4.624 | 36.274 | 2.706 | 25.531 | 4.965 | 11.535 | 1.318 | 4.854 | 40.679 |
EPE total |
EPE low-detail |
EPE high-detail |
EPE matched |
EPE unmatched |
EPE rigid |
EPE non-rigid |
EPE not sky |
EPE sky |
EPE s0-10 |
EPE s10-40 |
EPE s40+ |
0.475 | 0.437 | 6.438 | 0.380 | 4.388 | 0.203 | 2.529 | 0.476 | 0.457 | 0.129 | 0.375 | 3.639 |
Fl total |
Fl low-detail |
Fl high-detail |
Fl matched |
Fl unmatched |
Fl rigid |
Fl non-rigid |
Fl not sky |
Fl sky |
Fl s0-10 |
Fl s10-40 |
Fl s40+ |
1.621 | 1.451 | 28.389 | 1.267 | 16.273 | 0.821 | 7.676 | 1.667 | 0.932 | 0.382 | 2.588 | 9.376 |
WAUC total |
WAUC low-detail |
WAUC high-detail |
WAUC matched |
WAUC unmatched |
WAUC rigid |
WAUC non-rigid |
WAUC not sky |
WAUC sky |
WAUC s0-10 |
WAUC s10-40 |
WAUC s40+ |
92.720 | 93.026 | 44.474 | 93.393 | 64.892 | 95.184 | 74.075 | 93.035 | 87.929 | 96.727 | 91.541 | 62.385 |