Image dataset annotation for computer vision model

We are training a computer vision model and require a clean, carefully annotated training set based on our archive of source images. Volume — several thousand photographs that need to be marked with rectangular frames and polygonal masks, and for each object assign a class label according to a predefined taxonomy. We will provide written annotation instructions with examples, edge cases and rules for overlapping, cropped and overlaid objects, and expect the contractor to follow it precisely, not improvise. Uniformity across the entire set is more important to us than the speed of a single batch.

The result must be delivered in a standard annotation format so it integrates into our training pipeline without rework: one file per image, correct coordinates, stable class identifiers, no empty or orphaned records. First we want a small pilot batch, about one hundred images, to verify the taxonomy, agree on disputed cases and fix the instruction before starting the main volume. After that we will accept the rest in weekly deliveries. Please preserve folder structure and file naming exactly as specified: any manual renaming breaks our automatic import.

Quality control is part of the work, not a nice add-on. We expect re-review of a sample from each batch, a short report on disputed labels and prompt corrections to our comments. Confidentiality is mandatory: images remain in agreed storage and are not used anywhere else.

— Frames and polygonal masks, unified taxonomy
— Delivery in standard annotation format, one file per image
— Pilot batch, then weekly deliveries with quality control report