7.7x performance regression in experimental API polygon-to-mask conversion

[kprokofi]

Problem

Loading COCO-format instance segmentation datasets with polygon annotations through the experimental API is ~7.7x slower than equivalent Ultralytics dataloader.

Pipeline	Time per sample (ms)	Relative
Ultralytics (baseline)	24.5	1.0x
Datumaro experimental API	189.6	7.7x

Bottleneck

cProfile breakdown per sample:

Function	Time (ms)	%
mask_converters.convert()	45	24%
polygons_to_instance_masks()	33	17%
numpy.stack()	15	8%
Polars collect() / with_columns()	45	24%

Location: datumaro/experimental/converters/mask_converters.py

Root cause

Ultralytics converts polygons to masks using a tight C loop via cv2.fillPoly (~1-2ms for 10 instances). Datumaro uses Polars DataFrame operations with Python-level iteration and repeated numpy.stack calls.

Reproduction

from datumaro.experimental import Dataset

ds = Dataset.from_coco("path/to/coco", subset="train")
ds = ds.select(["image", "instance_mask"])

for i in range(100):
    sample = ds[i]
    _ = sample["instance_mask"]  # triggers conversion

Dataset: any COCO instance segmentation with polygon annotations (e.g., WGISD, COCO val2017).

Expected

Polygon-to-mask conversion performance should match or exceed cv2.fillPoly baseline.

Environment

• datumaro 1.11.0
• Python 3.11/3.13
• Linux x86_64

[kprokofi]

datumaro_perf_report.py