sangramrout/AIC_Challenge

AIC task-board port detection dataset (sfp_port / sc_port)

YOLO-format object-detection dataset for the connector ports on the task board of the AI for Industry Challenge cable-insertion task (a UR5e inserts a cable plug into a randomised task-board port).

Used to fine-tune sangramrout/AIC_Challenge (YOLOv8m), the only learned component in our solution.

Contents

split	images	labels
train	11,452	11,452
val	905	905

• Format: standard YOLO — images/{train,val}/*.png with matching labels/{train,val}/*.txt (class cx cy w h, normalised).
• Resolution: native 1152 × 1024 (not downscaled — small ports lose their edges when the frame is resized, which measurably hurt detection).
• Filename convention: {episode}_{frame}_{camera}.png, camera ∈ {left, center, right}.

Classes

id	name	description
0	sfp_port	SFP cage opening on a NIC card (two per card, 21.8 mm apart)
1	sc_port	SC fibre port on an SC rail module

Provenance

Fully synthetic. Frames are rendered from the challenge's Gazebo simulation via the three fixed scene cameras (left / center / right) across randomised task-board layouts — board pose, NIC-card rail slide, and SC rail slide all vary. No real photographs, and no human-annotated labels: boxes are derived from simulation ground truth, so they are exact.

Usage

pip install ultralytics

from ultralytics import YOLO

YOLO("yolov8m.pt").train(data="data.yaml", epochs=50, imgsz=1024)

data.yaml uses path: ., so it resolves relative to the dataset root — no editing needed.

Known pitfall — gate before you average

A detector trained on this data fires on every port on the board, not just the one you are targeting. The challenge's eval scene spawns three SC modules and multiple NIC cards, so taking a median over all detections of a class averages across distractor modules — we measured a 58 mm port-position error doing exactly that.

Select the target module first (gating by board-frame Y works well), then take the robust median of the survivors. That took our SC port lock from 58 mm → 2 mm with no change to the detector itself. The labels here are exact; the failure is entirely in how detections are aggregated.

Links

• Model: https://huggingface.co/sangramrout/AIC_Challenge
• Code: https://github.com/skr3178/aic