multi-data-gen / README.md
Lightning2030's picture
Initial dataset
53fab25
|
Raw History Blame Contribute Delete
2.36 kB
metadata
pretty_name: Multifaceted Generative Data Augmentation
task_categories:
  - object-detection
  - image-segmentation
size_categories:
  - 10K<n<100K
tags:
  - synthetic
  - pascal-voc
  - voc20

Multifaceted Generative Data Augmentation

This dataset contains 35,000 synthetic images with detection labels and 203,068 instance masks for the Pascal VOC 20 categories. Seven sources contain 5,000 images each.

Source Images Instances
test 5,000 30,763
test_nocomp 5,000 34,300
test_nosingle 5,000 40,842
test_novlmdata 5,000 28,358
test_onlycomp 5,000 26,692
test_onlysingle 5,000 12,548
test_onlyvlm 5,000 29,565

The sources represent the full single-class, composite, and VLM-guided generation method (test), each individual branch (test_onlysingle, test_onlycomp, test_onlyvlm), and the three two-branch combinations (test_nocomp, test_nosingle, test_novlmdata). Counts describe the complete generated collection.

Files

Each data/<source>.zip is paired with annotations/<source>.jsonl. Each JSONL contains exactly 5,000 records, one per image. Paths in image, label, and mask refer to members inside the named ZIP archive. Extracting a ZIP produces <source>/candidate/, <source>/labels/, and <source>/masks/<image_id>/.

A record contains id, archive, image, label, width, height, image_sha256, and instances. Each instance contains id (zero-based within its image), category_id, bbox_xyxy, mask, area, and mask_sha256. Bounding boxes use pixel coordinates [x_min, y_min, x_max, y_max]; mask area is the number of foreground pixels. PNG masks use values 0 and 255. TXT labels use normalized YOLO coordinates: category_id center_x center_y width height. Category IDs range from 0 to 19, as listed in annotations/categories.json.

manifest.json lists the data sources and checksums. files.jsonl.gz lists every image, label, and mask with its archive, byte size, and SHA256. SHA256SUMS contains repository file checksums. Summary counts and structural checks are available in annotations/summary.json, annotations/counts.csv, and annotations/validation.json.

Annotations are synthetic. Structural consistency checks do not measure accuracy against human reference annotations.