Checklist NEOAI 2026 Day 1 · Task 1
Terminal Cuties Segmentation
Train a segmentation model from scratch that separates cats and dogs from the background in images rendered as coloured ASCII art.
The task
The images are photographs of cats and dogs that were converted into coloured ASCII art ("ANSI TrueColor") and rendered back to PNG. The converter used a fixed palette of fill characters (" ", ".", ";", "c", "o", "P", "O", "?", "@", "■") and edge characters ("|", "-", "\", "/"); every character occupies an exact 10×10 pixel block, and each image is 128×128 characters, i.e. 1280×1280 pixels. The pets are surrounded by console clutter and ASCII-art background.
For every test image the contestant must predict a binary mask of the animal (character-level precision is sufficient, since every character is a 10×10 block). The exact font file used for rendering is supplied.
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
train.csv(id, img path, label mask path),test.csv(id, img path), rendered 1280×1280 PNG images, binary PNG masks with values 0 (background) and 255 (pet) for the training images, and font.ttf.- You submit
- CSV with columns id,
mask_rle: one run-length-encoded binary mask per test image (an empty string for an empty mask). The organisers' baseline writes a gzip-compressed CSV. - Scoring
- Dice coefficient (higher is better).
- Rules
- Pretrained models are not allowed; predefined architectures may be used if trained from scratch on the provided data.
- External data are prohibited.
- General NEOAI 2026 rules apply (see the rules page of the main competition).
- Format
- NEOAI 2026 main round, Day 1 (3 May 2026, 07:00–13:00 UTC per the main-competition schedule; the task competition closed at 13:00 UTC). Individual.