Complete code
qijiyaolinZGH opened this issue · comments
When will the Complete code be released?
The expected code release is object stage and fine stage optimization.
The input for each stage can be generated from existing segmentation/clip search on model database, I might not put all the dependency (as well as the whole model database) since some part of the pipeline is not public available. But I will update some more instructions on the code later next month.
I'm still sorting out the model database, does the object stage include the room stage as well? Any chance this gets released?
I haven't release the room stage yet, might be after sig asia.
When I was reproducing the object stage, I encountered the following issues: I reconstructed the point cloud using semantic and depth maps, and estimated the initial coordinates for each piece of furniture separately. However, the coordinates I estimated differ significantly from the ones you provided in "obj_stage_config.json". Here is the reconstructed point cloud I obtained. Could you help me analyze if I made a mistake in any step of the reproduction process?
I can send you the point cloud I predict to you about this weekend.
You might need to double check if your fov is correct while generate the pc. Also you might need to do a transform to the point cloud based on the camera to world or the angle of floor if I remember correctly.
I believe a transform based on the camera should be enough( from the output from perspective field, which is in a txt file includes camera fov/roll/pitch
I will need to double check later. (on a trip for siggraph asia now)
How do you normalize depth?Max-min normalization?
Can you provide your code for generating the point cloud?
Sure I will update the code for depth and depth point cloud later. The depth map is normalized with maximum depth as 1
(btw open a new github issue for these question in case I missed)