search

vinvino02 / GLPDepth

GLPDepth PyTorch Implementation: Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth
191 stars 44 forks source link

why do you divide depth by 1000 #28

Open dorohedorohe opened 2 years ago

dorohedorohe commented 2 years ago

As I understand you make the depth to.Tensor() so values are(0,1) and then you divide by 1000. So the gt has values ranging from(0,0.001)??

puyiwen commented 2 years ago

Because the gt depth is a 16-bit depth map stored in millimeters, so you need to divide by 1000 to get a depth map in meters. Like nyu_depth_v2, the values of gt ranging from (0-10).

sayakpaul commented 1 year ago

Then why 256 for the KITTI dataset?

cvrookieytd commented 1 year ago

please，Is the depth map generated by this network in meters?

Irish-kw commented 1 year ago

I found there are multiply 256 for kitti predict, 1000 for nyuv2 predict in the train.py (args.save_result) so divide 256 and 1000 is correct, may be just for save image with uint16