Detecting Chessboards

My arxiv frontpage found another fun article about a new dataset. The article in question is called "ChessVision -- A Dataset for Logically Coherent Multi-label Classification" and it introduces a dataset for detecting the state of a chessboard from a photo. The project repository can be found here.

Here are some examples from the dataset, as mentioned in the paper.

Images in the dataset with a corresponding chessboard state.

The 18.5Gb dataset has 200K+ of these images and it was able to generate this dataset by leveraging Blender. This is an open source tool for 3D modelling and animation which also comes with a neat Python integration. A bunch of computer vision datasets use it under the hood (another example here) and it makes sense that it's being used here as well. It took about 4 days to generate these images on 32 core CPU machine.

Another figure from the paper that shows the base image as well as other features.

The paper also lists some models with benchmarks, nit I do wonder how the dataset will be able to match with reality. The generated images might not resemble chess images from a tournament or a dark room, so I fear the application might be limited. Still, it's neat to see Blender being re-used for CV.