FGAHOI: Fine-Grained Anchors for Human-Object Interaction Detection
Human-Object Interaction (HOI), as an important problem in computer vision, requires locating the human-object pair and identifying the interactive relationships between them. The HOI instance has a greater span in spatial, scale, and task than the individual object instance, making its detection more susceptible to noisy backgrounds. To alleviate the disturbance of noisy backgrounds on HOI detection, it is necessary to consider the input image information to generate fine-grained anchors which are then leveraged to guide the detection of HOI instances. However, it has the following challenges. ๐) how to extract pivotal features from the images with complex background information is still an open question. ๐๐) how to semantically align the extracted features and query embeddings is also a difficult issue. In this paper, a novel end-to-end transformer-based framework (FGAHOI) is proposed to alleviate the above problems. FGAHOI comprises three dedicated components namely, multi-scale sampling (MSS), hierarchical spatial-aware merging (HSAM) and task-aware merging mechanism (TAM). MSS extracts features of humans, objects and interaction areas from noisy backgrounds for HOI instances of various scales. HSAM and TAM semantically align and merge the extracted features and query embeddings in the hierarchical spatial and task perspectives in turn. In the meanwhile, a novel training strategy Stage-wise Training Strategy is designed to reduce the training pressure caused by overly complex tasks done by FGAHOI. In addition, we propose two ways to measure the difficulty of HOI detection and a novel dataset, ๐.๐., HOI-SDC for the two challenges (Uneven Distributed Area in Human-Object Pairs and Long Distance Visual Modeling of Human-Object Pairs) of HOI instances detection. Experiments are conducted on three benchmarks: HICO-DET, HOI-SDC and V-COCO. Our model outperforms the state-of-the-art HOI detection methods, and the extensive ablations reveal the merits of our proposed contribution.
We test our models under python=3.8, pytorch=1.10.0, cuda=11.3. Other versions might be available as well.
conda create -n FGAHOI python =3.8 pip
conda activate FGAHOI
conda install pytorch==1.9.1 torchvision==0.10.1 torchaudio==0.9.1 cudatoolkit=11.3 -c pytorch -c conda-forge
pip install -r requirements.txt- Compiling CUDA operators
cd ./models/dab_deformable_detr/ops
sh ./make.sh
# test
python test.py- The details of argument can be changed in
main.py
Please follow the HICO-DET dataset preparation of GGNet.
After preparation, the data/hico_20160224_det folder as follows:
data
โโโ hico_20160224_det
| โโโ images
| | โโโ test2015
| | โโโ train2015
| โโโ annotations
| โโโ anno_list.json
| โโโ corre_hico.npy
| โโโ file_name_to_obj_cat.json
| โโโ hoi_id_to_num.json
| โโโ hoi_list_new.json
| โโโ test_hico.json
| โโโ trainval_hico.jsonPlease follow the installation of V-COCO.
For evaluation, please put vcoco_test.ids and vcoco_test.json into data/v-coco/data folder.
After preparation, the data/v-coco folder as follows:
data
โโโ v-coco
| โโโ prior.pickle
| โโโ images
| | โโโ train2014
| | โโโ val2014
| โโโ data
| | โโโ instances_vcoco_all_2014.json
| | โโโ vcoco_test.ids
| | โโโ vcoco_test.json
| โโโ annotations
| โโโ corre_vcoco.npy
| โโโ test_vcoco.json
| โโโ trainval_vcoco.jsonAfter preparation, the data/SDC folder as follows:
data
โโโ SDC
| โโโ JPGImages
| | โโโ image
| โโโ annotations
| โโโ train_annotation.json
| โโโ test_annotation.json
| โโโ train_split.txt
| โโโ test_split.txtMore details will come soon!
We currently provide results on HICO-DET.
| Model | Full (def) | Rare (def) | None-Rare (def) | Full (ko) | Rare (ko) | None-Rare (ko) | Weight |
|---|---|---|---|---|---|---|---|
| Swin-Tiny | 29.94 | 22.24 | 32.24 | 32.48 | 24.16 | 34.97 | Tiny_weight |
| Swin-Large*+ | 37.18 | 30.71 | 39.11 | 38.93 | 31.93 | 41.02 | Large_weight |
- Training FGAHOI with Swin-tiny from scratch.
stage base
python -m torch.distributed.launch \
--nproc_per_node=8 \
--use_env main.py \
--backbone swin_tiny \
--pretrained params/swin_tiny_patch4_window7_224.pth \
--dataset_file hico \
--num_verb_classes 117 \
--num_obj_classes 80 \
--output_dir logs/base \
--epochs 150 \
--lr_drop 120 \
--num_feature_levels 3 \
--num_queries 300 \
--merge \
--scale [1, 3, 5] \
--base \
--hoi_path data/hico_20160224_detstage hierarchical_merge
python -m torch.distributed.launch \
--nproc_per_node=8 \
--use_env main.py \
--backbone swin_tiny \
--pretrain_model_path "{Weights of the last stage}" \
--dataset_file hico \
--num_verb_classes 117 \
--num_obj_classes 80 \
--output_dir logs/hierarchical_merge \
--epochs 50 \
--lr_drop 40 \
--num_feature_levels 3 \
--num_queries 300 \
--merge \
--scale [1, 3, 5] \
--hierarchical_merge \
--hoi_path data/hico_20160224_det
stage hierarchical_merge and task_merge
python -m torch.distributed.launch \
--nproc_per_node=8 \
--use_env main.py \
--backbone swin_tiny \
--pretrain_model_path "{Weights of the last stage}" \
--dataset_file hico \
--num_verb_classes 117 \
--num_obj_classes 80 \
--output_dir logs/hierarchical_merge_and_task_merge \
--epochs 50 \
--lr_drop 40 \
--num_feature_levels 3 \
--num_queries 300 \
--merge \
--scale [1, 3, 5] \
--hierarchical_merge \
--task_merge \
--hoi_path data/hico_20160224_det- Evaluate FGAHOI with Swin-tiny from scratch.
python -m torch.distributed.launch \
--nproc_per_node=8 \
--use_env main.py \
--backbone swin_tiny \
--dataset_file hico \
--resume "{Weight of the model}"
--num_verb_classes 117 \
--num_obj_classes 80 \
--output_dir logs \
--epochs 150 \
--lr_drop 120 \
--num_feature_levels 3 \
--num_queries 300 \
--merge \
--scale [1, 3, 5] \
--hierarchical_merge \
--task_merge \
--eval \
--hoi_path data/hico_20160224_detThis repository is released under the Apache 2.0 license as found in the LICENSE file.
If you use FGAHOI, please consider citing:
@inproceedings{Ma2023FGAHOI,
title={FGAHOI: Fine-Grained Anchors forHuman-Object Interaction Detection},
author={Shuailei Ma and Yuefeng Wang and Shanze Wang and Ying Wei},
year={2023}
}
Should you have any question, please contact {[email protected]}
FGAHOI builds on previous works code base such as QAHOI, DAB-DETR. If you found FGAHOI useful please consider citing these works as well.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent