Updata.ca is a company from Québec that hired me for trying different architectures in their solution to classify waste recycling. I was in charge of testing instance segmentation models and they had a dataset of tens of millions images (artificially created). We trained models for days in 2 A100 (40Gb) cluster but results were not very satisfying due to a very difficult dataset: small inter-class dissimilarity and large intra-class variations we also had big occlusions in the images.
Due to NDA there are no images, sorry.
