Discord

Checklist IOAI 2025 Individual Contest · Task 6

Pixel Efficiency

For each animal photo, choose one rectangle covering at most 6.25% of the pixels so that a fixed CLIP model still classifies the species correctly.

  • Vision
  • Image masking for zero-shot classification

The task

Camera traps in remote wilderness send photos over very limited satellite links, so only the most useful pixels can be transmitted. You must decide which part of each 224 × 224 wildlife photo to keep.

For each image you return one rectangle, written as ((top, left), (bottom, right)) with the top-left corner inclusive and the bottom-right corner exclusive, covering at most 6.25% of the pixels. Everything outside it is painted black, and the result is classified zero-shot by the provided CLIP model.

An extra 'other' class is added to the species labels, so a mask that misses the animal cannot score by luck. You may work only with the provided CLIP model and dataset, and your code must process the 698 test images in under 8 minutes.

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
Datasets of 224×224 RGB animal photos with fields image, name (species; training set only) and idx: 700 training images and 698 hidden test images, split 30%/70% between Leaderboards A and B. A CLIP model is provided.
You submit
submission.jsonl with one rectangle per test image as ((top, left), (bottom, right)); the baseline writes one object per line with idx and coordinates keys. Submitted as submission.ipynb.
Scoring
Zero-shot classification accuracy of CLIP on the masked test images, with an added 'other' class; an invalid mask (wrong shape, more than 6.25% of pixels, etc.) counts as incorrect. The Scientific Committee's highest score (0.83) and baseline score (0.19) on Leaderboard B were used for score normalisation.
Rules
  • Keep at most 6.25% of each image's pixels, in a single rectangle
  • Work only with the provided CLIP model and dataset
  • Code must run in under 8 minutes for the 698 test images
  • At most 15 successful submissions
  • Test machine has no internet access
Format
Individual Contest, Day 2 (6 August 2025); six hours for the day's three tasks; the results page lists it as 'Pixel Parsimony Challenge'

Details

Year
2025, Beijing, China
Round
Individual Contest · Task 6
Language
English
License
CC BY 4.0, as stated by the source