GETTING_STARTED.md

Getting Started (Graph Generation)

Task-specific FCs

To run the graph generation baseline using task-specific FCs, run the script:

python run_easg.py annts_in_new_format data your_output_folder The data folder should contain

• Verb features
• RoI object features for train set
• RoI object features for validation set

This script also relies on the annotations in the format adjusted for SGG purposes stored in easg-generation/annts_in_new_format/.

The code for extracting object and verb features is provided in easg-generation/utils/. Verb features are based on Ego4D SlowFast features. Object features are based on detections of Faster-RCNN model, trained on our Ego4D-EASG dataset

STTran

This baseline builds on the original STTran repository. Please follow the instructions for requirements or other preparation.

First, download the Faster-RCNN checkpoint that is pre-trained on EASG and put it under fasterRCNN/models/res101/easg. Then, run the graph generation baseline with STTran as follows:

python train_with_EASG.py -mode $MODE -data_path $DATAPATH

MODE: edgecls, sgcls, or easgcls