search

DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images
MIT License
338 stars 97 forks source link

Weired ICP performance #59

Closed pengsida closed 4 years ago

pengsida commented 4 years ago

Hi, Your work is very very awesome, but I find something that confused me, which is that ICP improves the performance so much. The approach described in your paper follows a detector. Once the detection result is wrong, the initial pose could be far away from the ground truth pose, which makes the ICP hard to reach the optimal point. One plausible reason is that ICP can handle most poses whose 2d coordinates are within the image. Please tell me why ICP performs so good, since I am so curious.

MartinSmeyer commented 4 years ago

in the paper we use visual surface discrepancy (VSD) as a metric. VSD is very sensitive to offsets along the camera axis. While the orientation and 2D translation estimate are often correctly predicted using RGB data, the distance estimate using only RGB is often a bit inaccurate. A simple depth-based ICP along the camera thus lets the metrics increase a lot.

pengsida commented 4 years ago

Thanks for clarification!