Navigation Menu

Skip to content


Repository files navigation

ICCV'21 Context-aware Scene Graph Generation with Seq2Seq Transformers

Authors: Yichao Lu*, Himanshu Rai*, Cheng Chang*, Boris Knyazev†, Guangwei Yu, Shashank Shekhar†, Graham W. Taylor†, Maksims Volkovs

  • * Denotes equal contribution
  • † University of Guelph / Vector Institute

Prerequisites and Environment

  • pytorch-gpu 1.13.1
  • numpy 1.16.0
  • tqdm

All experiments were conducted on a 20-core Intel(R) Xeon(R) CPU E5-2630 v4 @2.20GHz and 4 NVIDIA V100 GPUs with 32GB GPU memory.


Visual Genome

Download it here. Unzip it under the data folder. You should see a vg folder unzipped there. It contains .json annotations that suit the dataloader used in this repo.

Visual Relation Detection

See Images:VRD


Visual Genome

Create a folder for all images:

# ROOT=path/to/cloned/repository
cd $ROOT/data/vg
mkdir VG_100K

Download Visual Genome images from the official page. Unzip all images (part 1 and part 2) into VG_100K/. There should be a total of 108249 files.

Visual Relation Detection

Create the vrd folder under data:

# ROOT=path/to/cloned/repository
cd $ROOT/data/vrd

Download the original annotation json files from here and unzip here. The images can be downloaded from here. Unzip to create an sg_dataset folder in data/vrd. Next run the preprocessing scripts:

cd $ROOT
python tools/
python tools/ converts all non-jpg images (some images are in png or gif) to jpg, and renames them in the {:012d}.jpg format (e.g., "000000000001.jpg"). It also creates new relationship annotations other than the original ones. This is mostly to make things easier for the dataloader. The filename mapping from the original is stored in data/vrd/*_fname_mapping.json where "*" is either "train" or "val". creates object detection annotations from the new annotations generated above, which are required by the dataloader during training.

Pre-trained Object Detection Models

Download pre-trained object detection models here. Unzip it under the root directory. Note: We do not include code for training object detectors. Please refer to the "(Optional) Training Object Detection Models" section in Large-Scale-VRD.pytorch for this.

Directory Structure

The final directories should look like:

|-- data
|   |-- detections_train.json
|   |-- detections_val.json
|   |-- new_annotations_train.json
|   |-- new_annotations_val.json
|   |-- objects.json
|   |-- predicates.json
|-- evaluation
|-- output
|   |-- pair_predicate_dict.dat
|   |-- train_data.dat
|   |-- valid_data.dat

Evaluating Pre-trained Relationship Detection models

DO NOT CHANGE anything in the provided config files(configs/xx/xxxx.yaml) even if you want to test with less or more than 8 GPUs. Use the environment variable CUDA_VISIBLE_DEVICES to control how many and which GPUs to use. Remove the --multi-gpu-test for single-gpu inference.

Visual Genome

NOTE: May require at least 64GB RAM to evaluate on the Visual Genome test set

We use three evaluation metrics for Visual Genome:

  1. SGDET: predict all the three labels and two boxes
  2. SGCLS: predict subject, object and predicate labels given ground truth subject and object boxes
  3. PRDCLS: predict predicate labels given ground truth subject and object boxes and labels

Training Scene Graph Generation Models

With the following command lines, the training results (models and logs) should be in $ROOT/Outputs/xxx/ where xxx is the .yaml file name used in the command without the ".yaml" extension. If you want to test with your trained models, simply run the test commands described above by setting --load_ckpt as the path of your trained models.

Visual Relation Detection

To train our scene graph generation model on the VRD dataset, run


python --dataset vrd --num-encoder-layers 4 --num-decoder-layers 2 --nhead 4



mv prediction.txt evaluation/vrd/

cd evaluation/vrd

python prediction.txt

Visual Genome

To train our scene graph generation model on the VG dataset, download the json files from, put the extracted files under data and then run


python --dataset vg --num-encoder-layers 4 --num-decoder-layers 2 --nhead 4



mv prediction.txt evaluation/vg/

cd evaluation/vg

python prediction.txt


This repository uses code based on the ContrastiveLosses4VRD Ji Zhang, Neural-Motifs source code from Rowan Zellers.


If you find this code useful in your research, please cite the following paper:

  title={Context-aware Scene Graph Generation with Seq2Seq Transformers},
  author={Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W. Taylor, Maksims Volkovs},


Code for the ICCV'21 paper "Context-aware Scene Graph Generation with Seq2Seq Transformers"






No releases published


No packages published