tchauffi/chesssight-synthetic-40k

ChessSight synthetic boards (40k)

Photorealistic Blender renders of chess positions with exact labels: the position as FEN and an 8×8 grid, the four board corners in image pixels, and an occlusion-correct box per piece. Built to train board readers without hand labelling, and validated by transfer to real photographs.

Licence and attribution — read this first

CC BY-NC 4.0. Commercial use is not permitted.

About 60% of these images render the Staunton Chess Set by uppalong (Printables model 76438), which is licensed CC BY-NC 4.0. The renders are derivative works, so the dataset inherits both the attribution requirement and the NonCommercial term. The remaining ~40% use a procedural piece set, and the HDRI environments come from Poly Haven (CC0) — but the two are not separable in this release, because the piece set drawn per sample was not recorded at generation time.

If you need a permissively licensed subset, regenerate with the procedural provider only using the code linked below.

Contents

40,000 images at 640×640, from two generator runs:

source_run	images	difference
train5	20,000	board border always darker than the dark squares
train6_border	20,000	border tone sampled across the full range, with printed rank/file coordinates

That difference is not cosmetic. A model trained only on train5 learns "the playing area ends where the surface darkens", which points at the wrong edge on a white tournament border. Matched training runs at 4,800 images per arm measured corner error of 0.154 squares (bordered) against 0.234 (unbordered) on real photographs — the bordered arm also beat a 16,000-image unbordered run, so the border variation was worth more than tripling the data.

Fields

field	description
image	640×640 RGB render
semantic_map	640×640 single-channel label map: 0 background, 1–12 piece classes, 13 board surface
instance_map	640×640 single-channel: 0 background, 1..N one id per piece
id	"<source_run>/<sample id>"
source_run	train5 or train6_border
fen	full FEN; only the placement field carries information
grid	64 ints, row-major from a8. 0 empty, 1–6 white PNBRQK, 7–12 black
corners_px	four [x, y] points, clockwise from the a8 corner
pieces	per piece: class_id, square, bbox (xyxy), visible, on_board
width, height	image size

Boxes are modal, measured from the instance mask, so they describe what is actually visible. Amodal boxes would teach a detector to hallucinate pieces hidden behind others.

Both label maps are PNG-encoded losslessly — verified identical after a round trip, and every visible piece has pixels of its own class inside its box. They come from a Blender object-index pass, so occlusion is exact rather than approximated: a pawn behind a queen is genuinely cut, not drawn over. The board carries its own semantic label because segmenting the playing surface is a useful auxiliary task — it is what a corner detector is really looking for.

Positions come from real Lichess games mixed with random legal-looking placements, so piece counts are not biased toward the opening. Some samples show captured pieces standing beside the board — those carry on_board: false.

Splits

train / validation / test at 80/10/10, assigned by hashing each sample id (blake2b), so the division survives regeneration and is identical to the one used for every number quoted here.

Reference results

Trained on this data alone, evaluated zero-shot on the real ChessReD test split:

model	metric
RT-DETR piece detector	0.6425 mAP, 0.8875 mAP@50
Swin-T corner heatmap	0.128 squares mean corner error, geometry found on 306/306
full pipeline (detector + corners + orientation)	97.06% of squares, 31.05% of boards exactly right

Board orientation — which corner is a8 — is resolved geometrically from square colour parity plus where White's men are, at 99.0% on ChessReD.

Known limitations

• Synthetic only. No real photographs are included. Use ChessReD or your own images for validation; training on renders and validating on renders will overstate what you have.
• Pawn detection is the weak class (~0.49–0.53 mAP on real photos) — small, low-contrast, and heavily occluded at grazing angles.
• A slight scale bias remains. Corner predictions come back about 1.4% too large on real boards, down from 3.3% before the border variation was added.
• The piece-set provider is not recorded per sample (see the licence note).

Generation

Code, generator configs and training recipes: https://github.com/tchauffi/ChessSight

Blender 5.2 / Cycles, 48 samples per image, ~2.3 images/second on an RTX 5070 Ti. Randomised across camera pose, focal length, lighting, board and piece materials, table surfaces, backdrops and distractor objects.