search

JinglinZhan / Real-Aug

Real-Aug: Realistic Scene Synthesis for LiDAR Augmentation in 3D Object Detection
Apache License 2.0
40 stars 0 forks source link

about error tolerance in distance constraint #4

Open mincheoree opened 1 year ago

mincheoree commented 1 year ago

Thanks for your interesting paper. I have one question about error tolerance setting in your paper Section 3.2.2. image The author mentioned that delta (error tolerance) is set as L/2 of object bounding box, but I am not sure if this tolerance is big enough that any lidar points in the scene fall onto distance constraint, because the points in lidar scene is quite sparse. I think there might be situation where the algorithm cannot find single (x,y) location to fit in the distance constraint, especially if gt sampled bbox's x and y are large. May I ask if you implement any algorithm for outlier cases? Thanks again for your hard work.

JinglinZhan commented 1 year ago

Thanks for your interesting paper. I have one question about error tolerance setting in your paper Section 3.2.2. image The author mentioned that delta (error tolerance) is set as L/2 of object bounding box, but I am not sure if this tolerance is big enough that any lidar points in the scene fall onto distance constraint, because the points in lidar scene is quite sparse. I think there might be situation where the algorithm cannot find single (x,y) location to fit in the distance constraint, especially if gt sampled bbox's x and y are large. May I ask if you implement any algorithm for outlier cases? Thanks again for your hard work.

In our implement, we simply discard the objects when there is no reasonable place for insertion. Perhaps increasing the preset number of inserted samples can compensate for the problem you raised.

mincheoree commented 1 year ago

Thanks for your interesting paper. I have one question about error tolerance setting in your paper Section 3.2.2. image The author mentioned that delta (error tolerance) is set as L/2 of object bounding box, but I am not sure if this tolerance is big enough that any lidar points in the scene fall onto distance constraint, because the points in lidar scene is quite sparse. I think there might be situation where the algorithm cannot find single (x,y) location to fit in the distance constraint, especially if gt sampled bbox's x and y are large. May I ask if you implement any algorithm for outlier cases? Thanks again for your hard work.

In our implement, we simply discard the objects when there is no reasonable place for insertion. Perhaps increasing the preset number of inserted samples can compensate for the problem you raised.

Thanks for your answer! Well understood.