DEPRECATED

containers-v1-3d-bbox-to-pixels

Containers v1 - Get the 3D bboxes of items in renders and project them to pixel space

There are 3 unique SKUs. Each render contains objects of the same SKU, with different color schemes. The bounding box of each SKU at zero rotation/translation is hard-coded. Based on the pose of each object in the info file, we calculate the axis-aligned 3d bounding box of each object, then project them to pixels.

There is information about the renders in the info.json and metadata.jsonl files. However, some information is missing. This data is hard-coded within the script:

1. The bounding box of each SKU at zero rotation/translation
2. Camera information (sensor size, field of view)

10 Sample images are included in the repo.

Usage

The script uses hydra for config files. Pass the root data directory containing the data to be processed as shown below. All the files (rgb, info json, metadata) must be present directly
in the root data directory

python containers_3d_bbox_to_pixels.py dir_data=sample_data/

There is a threshold to filter out objects that are not visible in the image, based on the number of pixels in an objects mask. Lowering this number will increase the sensitivity and show more objects that are only partially visible:

python containers_3d_bbox_to_pixels.py threshold_visible_object=1000

To enable debug information, pass the argument hydra.verbose=true. Debug will print a lot of information -> details of the transforms of every "visible" object in every image.

python containers_3d_bbox_to_pixels.py hydra.verbose=true

Install

Requires Python 3.

pip install -r requirements.txt