SceneParser/SceneParser-bench

SceneParser: Hierarchical Scene Parsing for Visual Semantics Understanding

Dataset release. JSONL annotations for SceneParser-Bench.

SceneParser-Bench provides JSONL annotations for hierarchical scene parsing. Each sample links an RGB image to a structured scene -> object -> part -> affordance hierarchy for training and evaluating SceneParser.

Files

train.jsonl
val.jsonl

Dataset Statistics

SceneParser-Bench is constructed based on EgoObjects as a large-scale benchmark for hierarchical scene parsing. It contains 110K training images and 5K validation images with no image overlap, annotated with explicit hierarchies linking localized objects, semantic parts, and affordances.

Split	Images	Objects	Parts	Affordances
Train	110K	743K	1.1M	1.67M
Val	5K	33.9K	49.5K	75.9K

In total, SceneParser-Bench contains 1.74M valid object-part-affordance chain instances.

Data Format

Each JSONL record uses a relative image path:

datasets/EgoObjects/images/<image_name>.jpg

Minimal schema:

{
  "image_path": "datasets/EgoObjects/images/example.jpg",
  "objects": [
    {
      "phrase": "object name",
      "object_bbox": [x1, y1, x2, y2],
      "parts": [
        {
          "part_name": "part name",
          "part_bbox": [x1, y1, x2, y2],
          "affordances": [
            {
              "action": "action name",
              "affordance_bbox": [x1, y1, x2, y2],
              "sampled_points": [[x, y]]
            }
          ]
        }
      ]
    }
  ]
}

Image Preparation

To reproduce training or evaluation, download the EgoObjects images from the official release and place the combined image folder at:

datasets/EgoObjects/images

Official image sources:

https://github.com/facebookresearch/EgoObjects
https://ai.meta.com/datasets/egoobjects-downloads/

Required archives:

EgoObjectsV1_images.zip
images.zip

Usage

Place train.jsonl and val.jsonl under the datasets directory of the SceneParser code release:

SceneParser/datasets/train.jsonl
SceneParser/datasets/val.jsonl

Then follow the conversion, training, and evaluation instructions in the SceneParser repository.

License

The SceneParser-Bench JSONL annotations are released under CC BY-NC 4.0.

The underlying images are not redistributed in this release. Users must download EgoObjects images from the official source and comply with the EgoObjects license (MIT) and any related terms from the data provider.