Transparent object segmentation

Less Real is More.

A minimal paper website for a synthetic-to-real ratio sweep on YOLOv26m-seg. The story is simple: a small real-data anchor helps a mostly synthetic mix outperform pure synthetic and pure real training.

636 real test images

2,509 annotated instances

9 training ratios

0.787 best segmentation mAP@50

Source Code Implementation and evaluation scripts Paper arXiv Preprint and citation page Model Hugging Face Weights, demos, or model cards Datasets Dataset All 9 ratio points, test & original

Method

Same recipe, different mix.

The site keeps the method intentionally compact: the only variable is the synthetic-to-real ratio.

Keep training fixed

Backbone, resolution, optimizer, augmentation, and seed stay constant across runs.

Sweep the ratio

Nine mixes run from 100:0 synthetic to real all the way to 0:100.

Evaluate on one real set

A fixed held-out real test set makes the comparison directly interpretable.

100:0 95:5 90:10 75:25 50:50 25:75 10:90 5:95 0:100

Results

Interactive ratio sweep.

Toggle metrics, inspect the highlighted mix, and sort the summary table below.

Live chart

Segmentation mAP@50

The main segmentation score peaks at 90:10, with 50:50 as a close secondary high point.

Segmentation view

Summary table

Sortable comparison across all ratios.

Click a header to sort. The table includes both segmentation and runtime metrics.

Ratio	Seg mAP@50	Seg mAP@50-95	Box mAP@50	Box mAP@50-95	Precision	Recall	Speed ms/image

Qualitative examples

Click a figure to inspect it.

These examples summarize the paper's visual story: clutter robustness, orientation generalization, occlusion handling, and better background recall with more real data.