Prepare Datasets for Training and Evaluating U2SEG

A dataset can be used by accessing DatasetCatalog for its data, or MetadataCatalog for its metadata (class names, etc). This document explains how to setup the builtin datasets so they can be used by the above APIs. Use Custom Datasets gives a deeper dive on how to use DatasetCatalog and MetadataCatalog, and how to add new datasets to them.

U2Seg has builtin support for a few datasets. The datasets are assumed to be placed under datasets directory.

All dataset annotation files should be converted to the COCO format.

Expected dataset structure for COCO:

├── coco
│   ├── annotations
│   │   ├── instances_val2017.json
│   │   └── instances_train2017.json
|   |   └── ......
│   ├── train2017
│   │   └── 000000000009.jpg
|   |   └── 000000000025.jpg
|   |   └── ......
│   ├── val2017
│   │   └── 000000000139.jpg
|   |   └── 000000000285.jpg
|   |   └── ......
├── datasets
|   |panoptic_anns
|   │   ├── panoptic_train2017.json
|   │   ├── panoptic_val2017.json
|   │   ├── panoptic_train2017
|   │   │   └── 000000000009.jpg
|   |   |   └── 000000000025.jpg
|   |   |   └── ......
|   │   ├── panoptic_val2017
|   │   │   └── 000000000139.jpg
|   |   |   └── 000000000285.jpg
|   |   |   └── ......

To generate the pseudo panoptic annotations for U2Seg training, we provide pre-generated anntation file at click, then put it in the structure of:

U2Seg-main/datasets
├── datasets
|   └── ......
├── prepare_ours
│   ├── u2seg_annotations
│   │   └── ins_annotations
|   |   └── semantic_annotations
|   |   └── panoptic_annotations (this will be generated using our code)

After you structure the file as above, you can generate the panoptic annotations (with 800 cluster centroid) by:

python ./datasets/prepare_ours/generate_pseudo_panoptic.py --class_num 800 --split train

This will automatically generate cocotrain_800 and cocotrain_800.json at the /prepare_ours/panoptic_annotationsfolder.

After that, run:

python ./datasets/prepare_ours/prepare_stuff_panoptic_fpn.py --cluster_num '800'

to generate the stuff annotations.

You can substitute the 800 mentioned above to any other number of classes used for clustering. (E.g. 300)

NOTE: ALL DATASETS FOLLOW THEIR ORIGINAL LICENSES.