harpreetsahota204/crop_query

Crop Query is a plugin for FiftyOne, an open-source tool used to organize and inspect image datasets. The plugin helps you find a specific object inside a large pile of unlabeled images when you only have a few example cutouts of that object. There is no model training and no fine-tuning involved, and the README says it does not require a GPU. The scenario it targets is common in computer vision work: you have thousands of pictures and a handful of cropped examples of something you care about, such as a bird's nest on a power line, a defect on a manufactured part, a sponsor logo in sports footage, or a specific vehicle in satellite images. Crop Query produces two outputs. The first is a ranked queue of images sorted by how closely they match the reference crops, with a tag applied to the top candidates. The second is a heatmap overlay shown on each image in the FiftyOne viewer that points to the region most likely to contain the object. The matching method is described step by step in the README. A pretrained vision model from the FiftyOne Model Zoo turns each reference crop into a numeric embedding. Each dataset image is cut into a grid of overlapping patches, and each patch is embedded with the same model. Cosine similarity is then computed between every patch and every template, and the highest similarity becomes the image's score. The per-patch similarities form the heatmap, with values below a chosen threshold hidden so only genuine matches stand out. The recommended starting model is clip-vit-base32-torch. The workflow inside FiftyOne is a sequence of small steps. You point the panel at a folder of template images, load a zoo model, then set the grid so that one patch roughly covers how much of an image the target object usually fills. You also pick the score field name, the tag to apply, and a cutoff for the heatmap. You can run the search on the entire dataset or on a current selection. The README lists possible use cases including industrial inspection, infrastructure monitoring on drone footage, camera trap filtering for wildlife, sponsor and team logo detection in broadcast archives, retail shelf checks, aerial and satellite imagery, and manufacturing quality checks.