Skip to content

Latest commit

 

History

History

demo

README.md

Webcam and Jupyter notebook demo

This folder contains a simple webcam demo where you can perform real-time SGG. First, make sure that you have downloaded or trained a model in SGDet mode, currently only sgdet is supported for demo. To run the demo you'll need to:

  1. Install the codebase, refer to INSTALL.md

  2. Train a SGDet model and save the results files in the ./checkpoints/ folder

  3. Get the path to the config file used for training, for instance ./configs/VG150/baseline/e2e_relation_X_101_32_8_FPN_1x.yaml

  4. Get the path of the dict file with class names, it should be under ./datasets/vg/ and be named something like VG-SGG-dicts.json

  5. Get the path of your trained weights, for instance ./checkpoints/upload_causal_motif_sgdet/model_0028000.pth

  6. OPTIONAL: if you want to spice it up, you can add a real-time tracker on top of the object detection to track relations between the same objects through time with boxmot (pip install boxmot). You can then activate it with argument --tracking.

  7. OPTIONAL: you can also configure the box and rel confidence threshold, by default I put 0.1 for the relation and 0.5 for the boxes, if no relation or too many relations are shown, try to adjust those thresholds with the arguments --rel_conf and --box_conf.

You can run the demo as follows:

python webcam_demo.py --classes PATH_TO_CLASSES_DICT_FILE.json --config YOUR_CONFIG_FILE_HERE.yml --weights YOUR_WEIGHTS_FILE.pth --tracking # only activate tracking if boxmot is installed