This is experimental work that never worked. I add some basic instructions, but do not expect any support nor answer to your questions.
- car/a1_coco_jsonToNegativesJson.m: It reads the COCO dataset and creates a JSON with all the images without cars on them. Note: I found in practice that COCO does not seem to label all cars, so unfortunately images with cars still appear.
- car/a2_valJsonsToImageValFolders.m: It creates the validation image folders for each one of the 3 datasets so the accuracy can be tested.
- car/a_jsonToTrainValJsons.m: Deprecated. Analog to
a2_valJsonsToImageValFolders.m, but only works with thecar-fusiondataset.
Even though this is unpublished work, it follows all the ideas from our main paper, Single-Network Whole-Body Pose Estimation. so please, cite it if this helps in your research.