Why stop at just office chairs? Why not detect all categorizable objects in every photo and rank based on that?
Could also crop just the object detection regions of each image, run those cropped images through CLIP/SigLIP, then UMAP and HDBSCAN to view a 2 or 3 dimensional latent space clustering of office chair types.. might reveal some info as to what kinds of chairs exist in what geographical regions. Could use a VLM to auto-tag each cluster given a couple images from each one. Could run PCA on the CLIP embeddings and have some sliders for each principal component.. maybe the first is chair color or size or whatever
Could also crop just the object detection regions of each image, run those cropped images through CLIP/SigLIP, then UMAP and HDBSCAN to view a 2 or 3 dimensional latent space clustering of office chair types.. might reveal some info as to what kinds of chairs exist in what geographical regions. Could use a VLM to auto-tag each cluster given a couple images from each one. Could run PCA on the CLIP embeddings and have some sliders for each principal component.. maybe the first is chair color or size or whatever
much data = much fun