Checklist IOAI 2025 Individual Contest · Task 2
Chicken Counting
Train a density-map model that counts Silkie chickens in farm photos, optionally building on a supplied pretrained feature extractor.
The task
Farmers and insurers need reliable counts of free-range Silkie chickens, for example to settle claims after losses to disease or predators. You are asked to build a counting model based on density estimation.
The training data has 100 RGB images (720 × 1280) with matching 180 × 320 density maps whose values sum to the number of chickens. Weights for the first four convolutional layers of a feature extractor are provided in base.pth. You may add your own density decoder on top of it, or build a model without it.
For the hidden validation and test sets (100 images each), your notebook must output non-negative 180 × 320 density maps. Predictions of the wrong shape or with negative values score zero. The statement calls this task a continuation of the at-home Satellite Weather Forecasting problem.
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
- 100 training images (RGB, 3×720×1280) with 180×320 density maps, plus pretrained feature-extractor weights (
base.pth); hidden validation (test_a) and test (test_b) sets of 100 images each. - You submit
submission.npzwith arrayspred_aandpred_b, each of shape 100×1×180×320 (100×180×320 is also accepted), holding non-negative density maps. The submitted notebook must include training and evaluation code.- Scoring
- Score = exp(-(1/n) Σ |
y_i- ŷ_i| /y_i), i.e. exp of minus the mean relative error of each image's total density. The Scientific Committee's highest score (0.89) and baseline score (0.71) were used for score normalisation. - Rules
- Output density maps must be 180×320 and non-negative, otherwise the score is zero
- Notebook must finish within 15 minutes on the test machine
- At most 15 successful submissions; up to 2 can be selected and the higher Leaderboard B score counts
- Only a notebook may be submitted (no self-made mounted datasets or .pth files)
- Test machine has no internet access; the provided
base.pthor other pretrained models that import offline may be used
- Format
- Individual Contest, Day 1 (5 August 2025); six hours for the day's three tasks