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/. (If the links are invalid, you can refer to their offical website: ImageNet-VidVRD dataset)

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	Recall@50	Recall@100
TRACE	30.6	19.3	24.6

VidVrd, detected box

Method	mAP	Recall@50	Recall@100
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}
}