search

BeileiCui / EndoDAC

[MICCAI'2024] EndoDAC: Efficient Adapting Foundation Model for Self-Supervised Depth Estimation from Any Endoscopic Camera
30 stars 2 forks source link

Pose-Intrinsics Net #9

Closed VincentItaliano closed 1 month ago

VincentItaliano commented 1 month ago

In the paper you say that the Pose-Intrinsics Net estimates the motion variation between adjacent images. By default we load 3 frames for each training image: the previous frame, the current frame, and the next frame. But in the code we only estimate the motion between the current and the next frame, but not between the previous frame and the current frame.

For exemple in file trainer_end_to_end.py, line 451 : 

for f_i in self.opt.frame_ids[1:]:

Do I need to modify this or have I misunderstood something?

BeileiCui commented 1 month ago

Hi, I think you misunderstood the code, The frame_ids is set to [0, -1,1] by default, so the inputs inputs_all = [pose_feats[f_i], pose_feats[0]] takes [0, -1] and [0, 1] by the for code, which means [current, previous] and [current next].

VincentItaliano commented 1 month ago

Hi, thank you for your reply. Indeed, I misread and was convinced I had read [-1, 0, 1] instead of [0, -1, 1] ...