Skip to content

Latest commit

 

History

History
182 lines (129 loc) · 9.81 KB

README.md

File metadata and controls

182 lines (129 loc) · 9.81 KB
Raw

Knowledge-Embedded Routing Network for Scene Graph Generation

Tianshui Chen*, Weihao Yu*, Riquan Chen, and Liang Lin, “Knowledge-Embedded Routing Network for Scene Graph Generation”, CVPR, 2019. (* co-first authors) [manuscript] (camera version will be available soon)

This repository contains trained models and PyTorch version code for the above paper, If the paper significantly inspires you, we request that you cite our work:

Bibtex

@inproceedings{chen2019knowledge,
  title={Knowledge-Embedded Routing Network for Scene Graph Generation},
  author={Chen, Tianshui and Yu, Weihao and Chen, Riquan and Lin, Liang},
  booktitle = "Conference on Computer Vision and Pattern Recognition",
  year={2019}
}

Setup

In our paper, our model's strong baseline model is SMN (Stacked Motif Networks) introduced by @rowanz et al. To compare these two models fairly, the PyTorch version code of our model is based on @rowanz's code neural-motifs. Thank @rowanz for sharing his nice code to research community.

  1. Install python3.6 and pytorch 3. I recommend the Anaconda distribution. To install PyTorch if you haven't already, use conda install pytorch=0.3.0 torchvision=0.2.0 cuda90 -c pytorch. We use TensorBoard to observe the results of validation dataset. If you want to use it in PyTorch, you should install TensorFlow and tensorboardX first. If you don't want to use TensorBaord, just not use the command -tb_log_dir.

  2. Update the config file with the dataset paths. Specifically:

    • Visual Genome (the VG_100K folder, image_data.json, VG-SGG.h5, and VG-SGG-dicts.json). See data/stanford_filtered/README.md for the steps to download these.
    • You'll also need to fix your PYTHONPATH: export PYTHONPATH=/home/yuweihao/exp/KERN

  3. Compile everything. run make in the main directory: this compiles the Bilinear Interpolation operation for the RoIs. Maybe you need to update your CUDA path in Makefile file.

  4. Pretrain VG detection. To compare our model with neural-motifs fairly, we just use their pretrained VG detection. You can download their pretrained detector checkpoint provided by @rowanz. You could also run ./scripts/pretrain_detector.sh to train detector by yourself. Note: You might have to modify the learning rate and batch size according to number and memory of GPU you have.

  5. Generate knowledge matrices: python prior_matrices/generate_knowledge.py, or download them from here: prior_matrices (Google Drive, OneDrive).

  6. Train our KERN model. There are three training phase. You need a GPU with 12G memory.

    • Train VG relationship predicate classification: run CUDA_VISIBLE_DEVICES=YOUR_GPU_NUM ./scripts/train_kern_predcls.sh This phase maybe last about 20-30 epochs.
    • Train scene graph classification: run CUDA_VISIBLE_DEVICES=YOUR_GPU_NUM ./scripts/train_kern_sgcls.sh. Before run this script, you need to modify the path name of best checkpoint you trained in precls phase: -ckpt checkpoints/kern_predcls/vgrel-YOUR_BEST_EPOCH_RNUM.tar. It lasts about 8-13 epochs, then you can decrease the learning rate to 1e-6 to further improve the performance. Like neural-motifs, we use only one trained checkpoint for both predcls and sgcls tasks. You can also download our checkpoint here: kern_sgcls_predcls.tar (Google Drive, OneDrive).
    • Refine for detection: run CUDA_VISIBLE_DEVICES=YOUR_GPU_NUM ./scripts/train_kern_sgdet.sh or download the checkpoint here: kern_sgdet.tar (Google Drive, OneDrive). If you find the validation performance plateaus, you could also decrease learning rate to 1e-6 to improve performance.

  7. Evaluate: refer to the scripts CUDA_VISIBLE_DEVICES=YOUR_GPU_NUM ./scripts/eval_kern_[predcls/sgcls/sgdet].sh. You can conveniently find all our checkpoints, evaluation caches and results in this folder KERN_Download (Google Drive, OneDrive).

Evaluation metrics

In validation/test dataset, assume there are images. For each image, a model generates top predicted relationship triplets. As for image , there are ground truth relationship triplets, where triplets are predicted successfully by the model. We can calculate:

For image , in its ground truth relationship triplets, there are ground truth triplets with relationship (Except , meaning no relationship. The number of relationship classes is , including no relationship), where triplets are predicted successfully by the model. In images of validation/test dataset, for relationship , there are images which contain at least one ground truth triplet with this relationship. The R@X of relationship can be calculated:

Then we can calculate:

Some results

relationship distribution of VG dataset Figure 1. The distribution of different relationships on the VG dataset. The training and test splits share similar distribution.


Ours_SMN Figure 2. The R@50 without constraint of our method and the SMN on the predicate classification task on the VG dataset.


Ours_SMN_Diff Figure 3. The R@50 absolute improvement of different relationships of our method to the SMN. The R@50 are computed without constraint.


Scatter Figure 4. The relation between the R@50 improvement and sample proportion on the predicate classification task on the VG dataset. The R@50 are computed without constraint.


Method SGGen SGCls PredCls Mean Relative
mR@50 mR@100 mR@50 mR@100 mR@50 mR@100 improvement
Constraint SMN 5.3 6.1 7.1 7.6 13.3 14.4 9.0
Ours 6.4 7.3 9.4 10.0 17.7 19.2 11.7 ↑ 30.0%
Unconstraint SMN 9.3 12.9 15.4 20.6 27.5 37.9 20.6
Ours 11.7 16.0 19.8 26.2 36.3 49.0 26.5 ↑ 28.6%
Table 1. Comparison of the mR@50 and mR@100 in % with and without constraint on the three tasks of the VG dataset.

Acknowledgement

Thank @rowanz for his generously releasing nice code neural-motifs.

Help

Feel free to open an issue if you encounter trouble getting it to work.