Although YCB Benchmarks has provided a series of mesh files, they still cannot be used directly in Gazebo because of the missing some of the physical properties.
I made SDF models for most YCB objects that can be used directly in Gazebo, and provided them with the following optimizations:
The YCB-Video dataset contributed by PoseCNN is based on the YCB dataset. 21 objects are selected, and photographed with an RGBD camera to make 92 videos. The entire data set contains 133827 frames.
The YCB-Video dataset is a little hard to download from Google Drive. Because it only consists of a super large compressed file (256GB) and does not support resumable transmission. Once the download fails, you will face an rate limitation (usually 24 hours) and have to download again from the first byte!😭
So I used sub-volume compression, each file is round 10GB in size with a verifiable hash. Although multi-threaded downloading is still not allowed, at least you can re-download from a sub-volume instead of the whole. Furthermore, I added text labels for YOLO training to each image under the data folder.
Continue reading