Target Adaptive Context Aggregation for Video Scene Graph Generation

This is a PyTorch implementation for Target Adaptive Context Aggregation for Video Scene Graph Generation.

Requirements

PyTorch >= 1.2 (Mine 1.7.1 (CUDA 10.1))
torchvision >= 0.4 (Mine 0.8.2 (CUDA 10.1))
cython
matplotlib
numpy
scipy
opencv
pyyaml
packaging
pycocotools
tensorboardX
tqdm
pillow
scikit-image
h5py
yacs
ninja
overrides
mmcv

Compilation

Compile the CUDA code in the Detectron submodule and in the repo:

# ROOT=path/to/cloned/repository
cd $ROOT/Detectron_pytorch/lib
sh make.sh
cd $ROOT/lib
sh make.sh

Data Preparation

Download Datasets

Download links: VidVRD and AG.

Create directories for datasets. The directories for ./data/ should look like:

|-- data
|   |-- ag
|   |-- vidvrd
|   |-- obj_embed

where ag and vidvrd are for AG and VidVRD datasets, and obj_embed is for GloVe, the weights of pre-trained word vectors. The final directories for GloVe should look like:

|-- obj_embed
|   |-- glove.6B.200d.pt
|   |-- glove.6B.300d.pt
|   |-- glove.6B.300d.txt
|   |-- glove.6B.200d.txt
|   |-- glove.6B.100d.txt
|   |-- glove.6B.50d.txt
|   |-- glove.6B.300d

AG

Put the .mp4 files into ./data/ag/videos/. Put the annotations into ./data/ag/annotations/.

The final directories for VidVRD dataset should look like:

|-- ag
|   |-- annotations
|   |   |-- object_classes.txt
|   |   |-- ...
|   |-- videos
|   |   |-- ....mp4
|   |-- Charades_annotations

VidVRD

Put the .mp4 files into ./data/vidvrd/videos/. Put the three documents test, train and videos from the vidvrd-annoataions into ./data/vidvrd/annotations/.

Download precomputed precomputed features, model and detected relations from here (or here). Extract features and models into ./data/vidvrd/.

The final directories for VidVRD dataset should look like:

|-- vidvrd
|   |-- annotations
|   |   |-- test
|   |   |-- train
|   |   |-- videos
|   |   |-- predicate.txt
|   |   |-- object.txt
|   |   |-- ...
|   |-- features
|   |   |-- relation
|   |   |-- traj_cls
|   |   |-- traj_cls_gt
|   |-- models
|   |   |-- baseline_setting.json
|   |   |-- ...
|   |-- videos
|   |   |-- ILSVRC2015_train_00005003.mp4
|   |   |-- ...

Change the format of annotations for AG and VidVRD

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

python tools/rename_ag.py

python tools/rename_vidvrd_anno.py

python tools/get_vidvrd_pretrained_rois.py --out_rpath pre_processed_boxes_gt_dense_more --rpath traj_cls_gt

python tools/get_vidvrd_pretrained_rois.py --out_rpath pre_processed_boxes_dense_more

Dump frames

Our ffmpeg version is 4.2.2-0york0~16.04 so using --ignore_editlist to avoid some frames being ignored. The jpg format saves the drive space.

Dump the annotated frames for AG and VidVRD.

python tools/dump_frames.py --ignore_editlist

python tools/dump_frames.py --ignore_editlist --video_dir data/vidvrd/videos --frame_dir data/vidvrd/frames --frame_list_file val_fname_list.json,train_fname_list.json --annotation_dir data/vidvrd/annotations --st_id 0

Dump the sampled high quality frames for AG and VidVRD.

python tools/dump_frames.py --frame_dir data/ag/sampled_frames --ignore_editlist --frames_store_type jpg --high_quality --sampled_frames

python tools/dump_frames.py --ignore_editlist --video_dir data/vidvrd/videos --frame_dir data/vidvrd/sampled_frames --frame_list_file val_fname_list.json,train_fname_list.json --annotation_dir data/vidvrd/annotations --frames_store_type jpg --high_quality --sampled_frames --st_id 0

If you want to dump all frames with jpg format.

python tools/dump_frames.py --all_frames --frame_dir data/ag/all_frames --ignore_editlist --frames_store_type jpg

Get classes in json format for AG

# ROOT=path/to/cloned/repository
cd $ROOT
python txt2json.py

Get Charades train/test split for AG

Download Charades annotations and extract the annotations into ./data/ag/Charades_annotations/. Then run,

# ROOT=path/to/cloned/repository
cd $ROOT
python tools/dataset_split.py

Pretrained Models

Download model weights from here.

pretrained object detection
TRACE trained on VidVRD in detection_models/vidvrd/trained_rel
TRACE trained on AG in detection_models/ag/trained_rel

Performance

VidVrd, gt box

Method	mAP	[email protected]	[email protected]
TRACE	30.6	19.3	24.6

VidVrd, detected box

Method	mAP	[email protected]	[email protected]
TRACE	16.3	9.2	11.2

AG, detected box

Training Relationship Detection Models

VidVRD

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

CUDA_VISIBLE_DEVICES=0 python tools/train_net_step_rel.py --dataset vidvrd --cfg configs/vidvrd/vidvrd_res101xi3d50_all_boxes_sample_train_flip_dc5_2d_new.yaml --nw 8 --use_tfboard --disp_interval 20 --o SGD --lr 0.025

AG

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

CUDA_VISIBLE_DEVICES=0 python tools/train_net_step_rel.py --dataset ag --cfg configs/ag/res101xi3d50_dc5_2d.yaml --nw 8 --use_tfboard --disp_interval 20 --o SGD --lr 0.01

Evaluating Relationship Detection Models

VidVRD

evaluation for gt boxes

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6,7 python tools/test_net_rel.py --dataset vidvrd --cfg configs/vidvrd/vidvrd_res101xi3d50_gt_boxes_dc5_2d_new.yaml --load_ckpt Outputs/vidvrd_res101xi3d50_all_boxes_sample_train_flip_dc5_2d_new/Aug01-16-20-06_gpuserver-11_step_with_prd_cls_v3/ckpt/model_step12999.pth --output_dir Outputs/vidvrd_new101 --do_val --multi-gpu-testing

python tools/transform_vidvrd_results.py --input_dir Outputs/vidvrd_new101 --output_dir Outputs/vidvrd_new101 --is_gt_traj

python tools/test_vidvrd.py --prediction Outputs/vidvrd_new101/baseline_relation_prediction.json --groundtruth data/vidvrd/annotations/test_gt.json

evaluation for detected boxes

CUDA_VISIBLE_DEVICES=1 python tools/test_net_rel.py --dataset vidvrd --cfg configs/vidvrd/vidvrd_res101xi3d50_pred_boxes_flip_dc5_2d_new.yaml --load_ckpt Outputs/vidvrd_res101xi3d50_all_boxes_sample_train_flip_dc5_2d_new/Aug01-16-20-06_gpuserver-11_step_with_prd_cls_v3/ckpt/model_step12999.pth --output_dir Outputs/vidvrd_new101_det2 --do_val

python tools/transform_vidvrd_results.py --input_dir Outputs/vidvrd_new101_det2 --output_dir Outputs/vidvrd_new101_det2

python tools/test_vidvrd.py --prediction Outputs/vidvrd_new101_det2/baseline_relation_prediction.json --groundtruth data/vidvrd/annotations/test_gt.json

AG

evaluation for detected boxes, Recalls (SGDet)

CUDA_VISIBLE_DEVICES=4 python tools/test_net_rel.py --dataset ag --cfg configs/ag/res101xi3d50_dc5_2d.yaml --load_ckpt Outputs/res101xi3d50_dc5_2d/Nov01-21-50-49_gpuserver-11_step_with_prd_cls_v3/ckpt/model_step177329.pth --output_dir Outputs/ag_val_101_ag_dc5_jin_map_new_infer_multiatten --do_val

#evaluation for detected boxes, mRecalls
python tools/visualize.py  --output_dir Outputs/ag_val_101_ag_dc5_jin_map_new_infer_multiatten --num 60000 --no_do_vis --rel_class_recall

evaluation for detected boxes, mAP_{rel}

CUDA_VISIBLE_DEVICES=4 python tools/test_net_rel.py --dataset ag --cfg configs/ag/res101xi3d50_dc5_2d.yaml --load_ckpt Outputs/res101xi3d50_dc5_2d/Nov01-21-50-49_gpuserver-11_step_with_prd_cls_v3/ckpt/model_step177329.pth --output_dir Outputs/ag_val_101_ag_dc5_jin_map_new_infer_multiatten --do_val --eva_map --topk 50

evaluation for gt boxes, Recalls (SGCls)

CUDA_VISIBLE_DEVICES=4 python tools/test_net_rel.py --dataset ag --cfg configs/ag/res101xi3d50_dc5_2d.yaml --load_ckpt Outputs/res101xi3d50_dc5_2d/Nov01-21-50-49_gpuserver-11_step_with_prd_cls_v3/ckpt/model_step177329.pth --output_dir Outputs/ag_val_101_ag_dc5_jin_map_new_infer_multiatten --do_val --use_gt_boxes

#evaluation for detected boxes, mRecalls
python tools/visualize.py  --output_dir Outputs/ag_val_101_ag_dc5_jin_map_new_infer_multiatten --num 60000 --no_do_vis --rel_class_recall

evaluation for gt boxes, gt object labels, Recalls (PredCls)

CUDA_VISIBLE_DEVICES=4 python tools/test_net_rel.py --dataset ag --cfg configs/ag/res101xi3d50_dc5_2d.yaml --load_ckpt Outputs/res101xi3d50_dc5_2d/Nov01-21-50-49_gpuserver-11_step_with_prd_cls_v3/ckpt/model_step177329.pth --output_dir Outputs/ag_val_101_ag_dc5_jin_map_new_infer_multiatten --do_val --use_gt_boxes --use_gt_labels

#evaluation for detected boxes, mRecalls
python tools/visualize.py  --output_dir Outputs/ag_val_101_ag_dc5_jin_map_new_infer_multiatten --num 60000 --no_do_vis --rel_class_recall

Hint

We apply the dilation convolution in I3D now, but observe a gridding effect in temporal feature maps.

Acknowledgements

This project is built on top of ContrastiveLosses4VRD, ActionGenome and VidVRD-helper. The corresponding papers are Graphical Contrastive Losses for Scene Graph Parsing, Action Genome: Actions as Compositions of Spatio-temporal Scene Graphs and Video Visual Relation Detection.

Citing

If you use this code in your research, please use the following BibTeX entry.

@inproceedings{Target_Adaptive_Context_Aggregation_for_Video_Scene_Graph_Generation,
  author    = {Yao Teng and
               Limin Wang and
               Zhifeng Li and
               Gangshan Wu},
  title     = {Target Adaptive Context Aggregation for Video Scene Graph Generation},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages     = {13688--13697},
  year      = {2021}
}

[ICCV 2021] Target Adaptive Context Aggregation for Video Scene Graph Generation

Related tags

Overview

Target Adaptive Context Aggregation for Video Scene Graph Generation

Requirements

Compilation

Data Preparation

Download Datasets

AG

VidVRD

Change the format of annotations for AG and VidVRD

Dump frames

Get classes in json format for AG

Get Charades train/test split for AG

Pretrained Models

Performance

Training Relationship Detection Models

VidVRD

AG

Evaluating Relationship Detection Models

VidVRD

AG

Hint

Acknowledgements

Citing

Owner

Multimedia Computing Group, Nanjing University

Code for You Only Cut Once: Boosting Data Augmentation with a Single Cut

Bayesian Optimization using GPflow

LF-YOLO (Lighter and Faster YOLO) is used to detect defect of X-ray weld image.

Code repo for realtime multi-person pose estimation in CVPR'17 (Oral)

An end-to-end PyTorch framework for image and video classification

RoboDesk A Multi-Task Reinforcement Learning Benchmark

labelpix is a graphical image labeling interface for drawing bounding boxes

Diabet Feature Engineering - Predict whether people have diabetes when their characteristics are specified

Research - dataset and code for 2016 paper Learning a Driving Simulator

This respository includes implementations on Manifoldron: Direct Space Partition via Manifold Discovery

BitPack is a practical tool to efficiently save ultra-low precision/mixed-precision quantized models.

BarcodeRattler - A Raspberry Pi Powered Barcode Reader to load a game on the Mister FPGA using MBC

Music Source Separation; Train & Eval & Inference piplines and pretrained models we used for 2021 ISMIR MDX Challenge.

Evaluation framework for testing segmentation networks in PyTorch

Unsupervised 3D Human Mesh Recovery from Noisy Point Clouds

A simple pygame dino game which can also be trained and played by a NEAT KI

Wide Residual Networks (WideResNets) in PyTorch

[ICCV 2021] Official PyTorch implementation for Deep Relational Metric Learning.

Contextual Attention Localization for Offline Handwritten Text Recognition

Code for Fold2Seq paper from ICML 2021