Code for reproducing the results in the following paper:
Learning to Detect Human-Object Interactions
Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, Jia Deng
IEEE Winter Conference on Applications of Computer Vision (WACV), 2018
Check out the project site for more details.
Please cite HO-RCNN if it helps your research:
@INPROCEEDINGS{chao:wacv2018,
author = {Yu-Wei Chao and Yunfan Liu and Xieyang Liu and Huayi Zeng and Jia Deng},
booktitle = {Proceedings of the IEEE Winter Conference on Applications of Computer Vision},
title = {Learning to Detect Human-Object Interactions},
year = {2018},
}
Some of the instructions in this README are modified from the fast-rcnn repo created by Ross Girshick.
This repo relies on two submodules (fast-rcnn
and caffe
), so make sure you clone with --recursive
:
git clone --recursive https://github.com/ywchao/ho-rcnn.git
- Evaluation
- Running Detection with a Trained Model
- Training a Model
- From R-CNN to Fast R-CNN
- Installation: Fast R-CNN
- Installation: MatCaffe
This demo runs the MATLAB evaluation script and replicates our results in the paper.
-
Download HICO-DET (7.5G):
./scripts/fetch_hico_det.sh
This will populate the
data
folder withhico_20160224_det
. -
Download pre-computed HOI detection on the HICO-DET test set (2.0G):
./scripts/fetch_hoi_detection.sh ./scripts/setup_symlinks_detection.sh
This will populate the
output
folder withprecomputed_hoi_detection
and set up a set of symlinks. -
Evaluating on 600 classes is tedious, so we use
parfor
to speed up. Uncomment and setpoolsize
inconfig.m
according to your need, or leave it commented out if you want MATLAB to set it automatically. -
Start MATLAB
matlab
underho-rcnn
. You should see the messageadded paths for the experiment!
followed by the MATLAB prompt>>
. -
Run
eval_run
. This will run the evaluation for the experiment HO+IP1 (conv)+S, under the Default and Known Object setting sequentially. The results will be printed and also saved under the folderevaluation/result
. -
You can run the evaluation for other experiments (which appear in the paper) by editing
eval_run.m
and rerunning it. For example, comment the line started withexp_name = 'rcnn_caffenet_ho_pconv_ip1_s';
and uncomment the line started withexp_name = 'rcnn_caffenet_ho_pconv_ip1';
.
This demo runs detection on the HICO-DET test set using a trained HO-RCNN model.
-
Obtain a trained model.
Option 1: Download pre-computed HO-RCNN models (4.0G).
./scripts/fetch_ho_rcnn_models.sh ./scripts/setup_symlinks_models.sh
This will populate the
output
folder withprecomputed_ho_rcnn_models
and set up a set of symlinks.Option 2: Train a Model yourself.
-
Download HICO-DET (7.5G) if you have not done so:
./scripts/fetch_hico_det.sh
This will populate the
data
folder withhico_20160224_det
. -
Download pre-computed Fast R-CNN object detection on the HICO-DET test set (37G):
./scripts/fetch_fast_rcnn_detection_test.sh
This will populate the
cache
folder withdet_base_caffenet/test2015
. -
Run HOI detection separately for 80 object classes. Take the experiment HO+IP1 (conv)+S for example. Run the 80 scripts under
experiments/scripts/test_rcnn_caffenet_ho_pconv_ip1_s
, which corresponds to 80 object classes of interest. Each script will load the detected objects of one class, classify the associated HOI classes, and generate an output file (e.g.detections_02.mat
) underoutput/ho_1_s/hico_det_test2015/rcnn_caffenet_pconv_ip_iter_150000
../experiments/scripts/test_rcnn_caffenet_ho_pconv_ip1_s/01_person.sh ./experiments/scripts/test_rcnn_caffenet_ho_pconv_ip1_s/02_bicycle.sh ... ./experiments/scripts/test_rcnn_caffenet_ho_pconv_ip1_s/80_toothbrush.sh
Warning: Finishing all 80 scripts may take very long. We run these scripts parallelly on multiple GPUs. If you find the required computation infeasible, you might want to consider a faster approach in From R-CNN to Fast R-CNN.
-
After verifying you have all 80 output files under
output/ho_1_s/hico_det_test2015/rcnn_caffenet_iter_150000
, you can run the evaluation following the previous section. Make sure to first remove the result files from evaluating pre-computed hoi detection, if any:rm -r evaluation/result
This demo trains a HO-RCNN model on the HICO-DET training set.
-
Install Fast R-CNN if you have not done so.
-
Download HICO-DET (7.5G) if you have not done so:
./scripts/fetch_hico_det.sh
This will populate the
data
folder withhico_20160224_det
. -
Download pre-computed Fast R-CNN object detection on the HICO-DET train set (145G):
./scripts/fetch_fast_rcnn_detection_train.sh
This will populate the
cache
folder withdet_base_caffenet/train2015
. -
Obtain the pre-trained ImageNet model.
Option 1: Download the post-surgery pre-trained ImageNet model. (recommended)
./scripts/fetch_post_surgery_imagenet_models.sh
This will populate the
data
folder withimagenet_models
.Option 2: Download the pre-trained ImageNet model and perform network surgery.
We use MatCaffe to perform network surgery, so you need to install MatCaffe first before running the commands below.
cd fast-rcnn ./data/scripts/fetch_imagenet_models.sh cd .. matlab -r "net_surgery; quit"
-
Remove the symlinks from previous sections, if any:
find output -type l -delete
-
Start training. Take the experiment HO+IP1 (conv)+S for example:
./experiments/scripts/rcnn_caffenet_ho_pconv_ip1_s.sh
The trained models will be saved in
output/ho_1_s/hico_det_train2015
and the log file will be saved inexperiments/logs
. -
Once the training is complete. You can run detection and run evaluation following the previous sections.
-
Note that HO-RCNN performs detection in the R-CNN style, i.e. proposals from the same image are independently processed by the network. This is not very efficient especially in the test time. One way to speed things up is to perform detection in the Fast R-CNN style, i.e. applying the convolutional layers on the whole image followed by extracting proposal specific features with RoI pooling. This has already been implemented and included in this repo.
-
We observed a 2-3x speedup on a full training run, and a 5-35x speedup on a full test run, where the amount of speedup is primarily determined by the network architecture. However, we also observed a slight decrease in mAP.
-
You only need to run one script to run both training and test. First, complete step 1 to 5 in Training a Model and step 1 to 4 in Running Detection with a Trained Model. After that, run the script under
experiments/scripts
with prefixfast_rcnn_
instead ofrcnn_
. Take the experiment HO+IP1 (conv)+S for example:./experiments/scripts/fast_rcnn_caffenet_ho_pconv_ip1_s.sh
The trained models will be saved in
output/ho_1_s/hico_det_train2015
but now with prefixfast_rcnn_
instead ofrcnn_
. The log file will be saved inexperiments/logs
as before. The test output will now be just a single filedetections.mat
underoutput/ho_1_s/hico_det_test2015/fast_rcnn_caffenet_pconv_ip_iter_150000
. -
To run evaluation, you only need to slightly edit
eval_run.m
and run it. Take the experiment HO+IP1 (conv)+S for example: comment the line started withexp_name = 'rcnn_caffenet_ho_pconv_ip1_s';
and uncomment the line started withexp_name = 'fast_rcnn_caffenet_ho_pconv_ip1_s';
.
The training and test of HO-RCNN is implemented in our own branch of Fast R-CNN.
To install Fast R-CNN for HO-RCNN, change the directory:
cd fast-rcnn
and go through the below steps in this README:
- Requirements: software
- Build the Cython modules
- Build Caffe and pycaffe
Note:
- We built Caffe with cuDNN v2 (cudnn-6.5-linux-x64-v2), which is the only cuDNN version the given branch supports.
- All our experiments are ran on the GeForce GTX TITAN X GPU.
MatCaffe is only needed if you want to run the network surgery yourself (see: Training a Model).
To install MatCaffe:
cd caffe
# Now follow the Caffe installation instructions here:
# http://caffe.berkeleyvision.org/installation.html
# If you're experienced with Caffe and have all of the requirements installed
# and your Makefile.config in place, then simply do:
make -j8 && make matcaffe