Discord

Checklist APOAI 2026 Contest · Task 4

Spot the Spots: Individual Identification of Leopard Cats

Identify which of 54 individual leopard cats appears in a camera-trap image, using a small set of permitted pretrained backbones.

  • Vision
  • Fine-grained image classification / individual re-identification

The task

Camera traps capture wildlife non-invasively; beyond species recognition, identifying individuals allows population and movement studies. Using a dataset from Professor Luo Shu-Jin's team at Peking University on leopard cats (Prionailurus bengalensis) near Beijing, the contestant builds a system that recognises individuals from their spot patterns, robust to day/night infrared lighting, clutter and pose.

The curated dataset has 1,090 images of 54 individuals (labels 0–53): 780 training, 155 validation (public leaderboard) and 155 test (private leaderboard). Labels are in CSV files (filename, label); validation and test images are only accessible through environment variables during the official submission run.

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
Training images (train/*.jpg) and train.csv; a dataset of permitted pretrained weights (ConvNeXt-Tiny, ResNet101 and Swin-T from torchvision, Swin-Base 384 from Hugging Face, YOLO11 n/s/m/x from ultralytics).
You submit
A notebook (training and prediction) producing submission.zip with validation_submission.csv and test_submission.csv (columns filename, label).
Scoring
Accuracy. Leaderboard A = validation, B = test. Baseline on B: 0.2580; scientific committee reference: 0.6580.
Rules
  • Only the listed pretrained weights may be used; any other pretrained weights may result in a score of zero.
  • Only the train split may be used for model development.
  • Training and inference within 25 minutes on a Tesla L20 GPU.
  • Self-contained, reproducible, no external API calls or internet access.
  • Evaluated in the 'ioai3.7' image (the only task using it).
Format
APOAI 2026 contest (first edition): a six-hour examination on Saturday 27 June 2026, taken online on the Bohrium platform with local invigilated sites; the four tasks were republished on Bohrium for practice from 15 June to 31 July 2026 (platform time UTC+8).

Details

Year
2026, Online
Round
Contest · Task 4
Language
English
License
Not stated by the source