Checklist NEOAI 2025 Day 3 · Task 7
Cuties Segmentation
Produce binary segmentation masks of cats and dogs using only a single pretrained CLIP model and 20 labelled validation images.
The task
The images show cats and dogs of different breeds; a list of 37 breed names present in the data is given (not every breed is guaranteed to appear in the validation data). The contestant must output a binary segmentation map of the animal for each of the 1,000 test images.
Only the supplied CLIP model (openai/clip-vit-base-patch16) may be used as a pretrained model. The baseline performs zero-shot breed classification with CLIP and then scores 16×16 patches against the predicted class name to obtain a heat map, which is thresholded.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- cuties/
val_imgsand cuties/val_masks(20 images with binary masks), cuties/test_imgs(1,000 images), the list of breed names, and the CLIP model. - You submit
- CSV with columns
img_idand mask, where mask is a base64-encoded PNG binary mask at the original image size. - Scoring
- Binary-mask IoU (intersection over union), as implemented in the notebook's binaryMaskIOU; the Kaggle metric is a custom template.
- Rules
- No pretrained models except the given CLIP.
- No external datasets.
- Any prompts for CLIP may be written; training on the validation data is allowed.
- General NEOAI 2025 rules apply: by default no pretrained models or additional data unless the statement allows them; no external human help; on-site participants get one NVIDIA V100, an LLM chat (GPT-4o or similar) and a whitelist of websites.
- Format
- NEOAI 2025 main round, Task 7, Day 3 (9 May 2025, 07:00–13:00 UTC). Individual; 20 submissions per day.