Introduction to the DeSOPE Dataset

This directory contains RGB image sequences, object pose annotations, camera intrinsics, object mesh models, and a utility script for visualizing object contours.

Dataset Structure

scene_1/
├── cam_K.txt
├── models/
│   ├── cheese_wafer_box_canonical.ply
│   ├── cheese_wafer_box_deform_mild.ply
│   └── ...
├── static/
│   ├── seq_1/
│   │   ├── pose.json
│   │   └── rgb/
│   │       ├── 000000.png
│   │       ├── 000002.png
│   │       └── ...
│   │   └── depth/
│   │       ├── 000000.png
│   │       ├── 000002.png
│   │       └── ...
│   │   └── masks/
│   │       ├── 000000
│   │       ├── 000002
│   │       └── ...
│   ├── seq_2/
│   │   ├── pose.json
│   │   └── rgb/
│   │       └── ...
│   └── ...
├── models_name.md
├── seq_img_obj.md

Main Files and Directories

• static/: contains different sequence folders named as seq_{seq id}.

• static/seq_{seq id}/rgb/: contains RGB images for the sequence. Image filenames use 6-digit IDs such as 000000.png.

• static/seq_{seq id}/pose.json: stores object poses for images in the sequence. Each pose entry contains an obj_id, cam_R_m2c, and cam_t_m2c.

• models/: contains object mesh models in PLY format. For example, models/1.ply corresponds to object ID 1.

• cam_K.txt: stores the camera intrinsic matrix.

• seq_img_obj.md: lists which objects have valid poses in each image of each sequence.

Data Index

Before running the visualization script, please check seq_img_obj.md and select the target entry from its table:

• seq id: the sequence ID, corresponding to static/seq_{seq id}
• img id: the image ID, corresponding to static/seq_{seq id}/rgb/{img id}.png
• obj id: the object ID with pose annotation in that image, corresponding to models/{obj id}.ply

Only the seq id, img id, and obj id combinations listed in seq_img_obj.md have valid poses and can be used for contour visualization.

Object Contour Visualization

visualize_object_contour.py renders the contour of a specified object using:

1. the RGB image from static/seq_{seq id}/rgb/{img id}.png,
2. the object pose from static/seq_{seq id}/pose.json,
3. the object mesh from models/{obj id}.ply, and
4. the camera intrinsics from cam_K.txt or cam_K.json.

Basic Usage

python visualize_object_contour.py --seq-id 1 --img-id 000000 --obj-id 1

You can also provide img id in numeric form:

python visualize_object_contour.py --seq-id 1 --img-id 0 --obj-id 1

The script automatically pads numeric image IDs to 6 digits. For example, 0 is parsed as 000000.

Specify Output Path

python visualize_object_contour.py \
  --seq-id 1 \
  --img-id 000000 \
  --obj-id 1 \
  --output outputs/seq1_img000000_obj1.png

If --output is not specified, the default output path is:

edge_seq{seq_id}_img{img_id}_obj{obj_id}.png

Common Options

python visualize_object_contour.py \
  --seq-id 1 \
  --img-id 000000 \
  --obj-id 1 \
  --edge-thickness 3 \
  --color 255 0 0 \
  --alpha 1.0

Option descriptions:

• --seq-id: sequence ID, required
• --img-id: image ID, required
• --obj-id: object ID, required
• --output: output image path, optional
• --edge-thickness: contour line width, default is 3
• --color R G B: contour color, default is red 255 0 0
• --alpha: contour overlay opacity in [0, 1], default is 1.0
• --root: dataset root directory, default is the current directory
• --camera: camera intrinsic file path; if not specified, the script uses cam_K.json first, otherwise cam_K.txt
• --occurrence: selects which instance to visualize when the same obj id appears multiple times in one image, default is 0

Example

Green semi-transparent contour:

python visualize_object_contour.py \
  --seq-id 1 \
  --img-id 000000 \
  --obj-id 1 \
  --color 0 255 0 \
  --alpha 0.8