Seeing the World through Any Eyes

TL;DR: EgoEye generates realistic egocentric videos from a single exocentric video by combining reward-guided egocentric reasoning, multi-stage motion alignment, and egocentric pretraining, while EgoScape provides a large-scale distortion-free paired exo-ego dataset and benchmarks for training and evaluation.

View Comparisons View Dataset Samples

Comparison on EgoEval

Comparison on EgoEval

Comparison on EgoCross

Comparison on EgoCross

Comparison on EgoWild

Comparison on EgoWild

More EgoEye Results on EgoWild

Click any result to switch. The player also advances automatically when the current video ends.

EgoScape Dataset Samples

The following eight paired exocentric-egocentric samples are selected from the first eight filenames in the sorted dataset list.

Dataset sample 1

20260107_0011_00000030_day_single_move_2

Dataset sample 2

20260107_0013_00000130_day_single_move_1

Dataset sample 3

20260107_0016_00000050_day_single_move_1

Dataset sample 4

20260107_0017_00000360_day_single_move_1

Dataset sample 5

20260107_0023_00000030_day_single_static_1

Dataset sample 6

20260107_0068_00000170_day_single_static_1

Dataset sample 7

20260107_0084_00000120_night_single_move_1

Dataset sample 8

20260108_0008_00000040_day_single_static_1