Re:Ask HN: Data augmentation tools for 3d object detection
by mathgaron on Friday December 07, 2018 @09:06AM
(#18627709)
Are you talking about a depth channel on an image plane? It also largely depend on the problem but here is a few tricks that helped me for my problems (object tracking).
- Generating synthetic data is powerful if you have the depth modality as it is easy to render. Also the real/synthetic domain gap is narrow compared to RGB. I consider it as data augmentation: you usually do many renders from a single 3D model.
- If you can somehow normalize the offset (e.g. compute normals) that can help. In my case I could offset the center of the object as 0 depth and it greatly help the network to converge.
- Classic augmentations like gaussian noise, gaussian blur and also downsampling the depth helps (apply these randomly).
As for tooling, I just use numpy/pytorch for most operations and OpenGL for renders.
- Generating synthetic data is powerful if you have the depth modality as it is easy to render. Also the real/synthetic domain gap is narrow compared to RGB. I consider it as data augmentation: you usually do many renders from a single 3D model.
- If you can somehow normalize the offset (e.g. compute normals) that can help. In my case I could offset the center of the object as 0 depth and it greatly help the network to converge.
- Classic augmentations like gaussian noise, gaussian blur and also downsampling the depth helps (apply these randomly).
As for tooling, I just use numpy/pytorch for most operations and OpenGL for renders.