Visual Manipulation Relationship Network

News! I have updated this package so that it can support pytorch 1.0 now. See branch pytorch1.0.

IMPORTANT: Please be informed that the performance of the pytorch1.0 version is not the ones in the paper. That is:

1. For VMRN: the performance is lower due to some unknown reasons.

2. For ROI-GD: the performance is much better.

I am trying to fix the bugs here. Before further updates, please use the pytorch0.4.0_oldversion branch to reproduce the performance in the paper.

Introduction

This package includes our latest proposed robotic grasping algorithms. Main framework is based on code of Faster RCNN (https://github.com/jwyang/faster-rcnn.pytorch).

Implemented Algorithms

I have included some well-known object detection algorithms (e.g. Faster-RCNN, SSD, and FPN), and developed a series of algorithms related to robotc grasping. The following figure provides an overview of this package.

In the future, I will continuously update this package so that it can contain latest SOTA algorithms and improve the performance of our own algorithms. Certainly, all my new works related to robotic grasping and computer vision will also be contained in this package.

I stopped maintain this package. Now I am transferring to MMDetection. And maybe not further in the future, I will make my mmdet-based library open-sourced, which is with higher performance.

Installation

Since we use torchvision.ops instead of all previous C expansions, all codes can run normally without any extra building. Follow training example to enjoy it.

Training Example

python main.py --dataset (DatasetName) --frame (AlgName) --net (BackboneName) --cuda
# like:
python main.py --dataset vmrdcompv1 --frame all_in_one --net res101 --cuda

Testing

python main.py --test --dataset (DatasetName) --frame (AlgName) --net (BackboneName) --cuda --checkpoint (PointNum) --checkepoch (EpochNum) --GPU (GpuNum, Default:0)
#like:
python main.py --test --dataset vmrdcompv1 --frame all_in_one --net res101 --cuda --checkpoint 1000 --checkepoch 1 --GPU 0

Performance

Some performance is shown below and it will be updated continuously.

mAP: mean Average Precision

mAP-G: mean Average Precision with Grasp Detection

Rel-IA: Image Accuracy of Relationship Detection

Performance Summary

Algorithm	Backbone	Training	Testing	mAP	mAP-G	Rel-IA
Faster R-CNN	ResNet-101	VOC2007trainval	VOC2007test	71.5	-	-
FPN	ResNet-101	VOC2007trainval	VOC2007test	73.9	-	-
ROI-GD	ResNet-101	VMRDtrainval	VMRDtest	94.5	75.7	-
F-VMRN	ResNet-101	VMRDtrainval	VMRDtest	95.6	-	64.7
F-VMRN	VGG-16	VMRDtrainval	VMRDtest	95.0	-	68.7

Testing Curves

Rel-IA of F-VMRN_VGG16	Rel-IA of F-VMRN_ResNet101

mAP of Faster-RCNN on VOC2007

TODO

• Performance of SSD, S-VMRN, All-In-One
• Testing Curve of Faster-RCNN, FPN, ROI-GD
• Complete VAM model
• COCO, VisualGenome Support
• Multi-GPU Support
• PyTorch 1.0 Support (see branch pytorch1.0)
• Python3 Support

Noteable Things

1. To train the network, you have to pre-download the pretrained models and put them in "data/pretrained_model" and name them the same as the usage in codes.
2. The training data should be placed or linked in "data".
3. The code will be improved continuously. Therefore, if you meet some problems, do not hesitate to contact me.
4. The included FPN and Focal Loss are uncompleted while Faster-RCNN and SSD can be used normally, though they are not the main contributions.
5. [UPDATE in 01/04/2020] I have re-written all codes so that they can be more readable and modularized. Therefore, if you want to obtain the performance claimed in the paper, you should check the configurations so that they are same as the claimed ones (e.g. data-augmentation methods).

Papers

1. Zhang, Hanbo, et al. "ROI-based Robotic Grasp Detection for Object Overlapping Scenes." arXiv preprint arXiv:1808.10313 (2018). To appear in IROS 2019.
2. Zhang, Hanbo, et al. "A Multi-task Convolutional Neural Network for Autonomous Robotic Grasping in Object Stacking Scenes." arXiv preprint arXiv:1809.07081 (2018). To appear in IROS 2019.
3. Zhang, Hanbo, et al. "Visual Manipulation Relationship Network for Autonomous Robotics." 2018 IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids). IEEE, 2018.
4. Zhang, Hanbo, et al. "A Real-time Robotic Grasp Approach with Oriented Anchor Box." IEEE Transactions on Systems, Man and Cybernetics: Systems. Online Early Access.
5. Zhou, Xinwen, et al. "Fully convolutional grasp detection network with oriented anchor box." 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018.
6. Ren, Shaoqing, et al. "Faster r-cnn: Towards real-time object detection with region proposal networks." Advances in neural information processing systems. 2015.
7. Liu, Wei, et al. "Ssd: Single shot multibox detector." European conference on computer vision. Springer, Cham, 2016.
8. Lin, Tsung-Yi, et al. "Focal loss for dense object detection." Proceedings of the IEEE international conference on computer vision. 2017.

Problem Shooting

1. [Solved] When setting batch_size of Faster RCNN to 1 and augmentation to True, I want to use SSD-like augmentation to generate more training data. However, it will cause NaN error.
2. There are some grasp and relation label errors in VMRD. However, I find that they do not affect the detection performance much. I will fix this problem as soon as possible.
3. This package only supports pytorch 0.4.0. When using 0.4.1, there will be segmentation fault (reasons not found).
4. When training VMRN, the batchsize should be at least 2. Otherwise the network will not reach convergence.
5. All vggXX_bns cannot work well.