The code is provided for training and testing VL-MPAG Net on Visual Genome and COCO-Stuff datasets.

Installation

The installation requirements are given in requirements.txt file.

Data prepration

Visual Genome dataset

Download the Visual Genome dataset images link.

The splits of the Visual Genome dataset used in our analysis can be accessed by use the following anonymous link

https://github.com/anonymous9039/SGL_data

COCO-stuff dataset

Downoad the COCO-stuff dataset from link

Training

The training instructions are provided for Visual Genome and COCO-stuff datasets.

Training on Visual Genome

The training code is provided for both Visual Genome Full Overlap(VGFO) and Visual Genome Partial Overlap (VGPO) setting.

For training the model on VG dataset run the following command.

python train_detector_vg.py

Training on COCO-stuff dataset

Run the following commands for training.

python train_detector_coco.py

You can find the hyperparameters of the model inside the training files.

Testing

Visual Genome

For testing the model on VG run the following command.

python test_detector_vg.py

COCO-stuff

python test_detector_coco.py