Fast3R than Fast3R .
Alternative attention
24 layers of [first attention between tokens within images, then global attention between all tokens]. No positional embedding for views except the first one, making this method permutation equivariant (this makes it unable to exploit sequences?).
Two heads
- Camera head is an MLP that predicts intrinsics and extrinsics for each view
- DPT for dense predictions – depthmap, pointmap, tracks (dense features that some downstream function turns into actual tracks)
Training
9 days in 64xA100, bfloat16