This is an tensorflow implementation of "CloudAAE: Learning 6D Object Pose Regression with On-line Data Synthesis on Point Clouds"
- log: directory to store log files during training.
- losses: loss functions for training.
- models: a python file defining model structure.
- object_model_tfrecord: full object models for data synthesizing and visualization purpose.
- tf_ops: tensorflow implementation of sampling operations (credit: Haoqiang Fan, Charles R. Qi).
- trained_network: a trained network.
- utils: utility files for defining model structure.
- ycb_video_data_tfRecords: synthetic training data and real test data for the YCB video dataset.
- evaluate_cloudAAE_ycbv.py: script for testing object 6d pose estimation with a trained network on test set in YCB video dataset.
- train_cloudAAE_ycbv.py: script for training a network on synthetic data for YCB objects.
Test a trained network
- Testing data in tfrecord format is available
- Download zip file
- Unzip and place all files in ycb_video_data_tfRecords/test_real/
After activate tensorflow
python evaluate_cloudAAE_ycbv.py --trained_model trained_network/20200908-204328/model.ckpt --batch_size 1 --target_cls 0
- --trained_model: directory to trained model (*.ckpt).
- --batch_size: 1.
- --target_class: target class for pose estimation.
- Translation prediction is in unit meter.
- Rotation prediction is in axis-angle format.
- If you turn on visualization with b_visual=True, you will see the following displays which are partially observed point cloud segments (red) overlaid with object model (green) with pose estimates. The reconstructed point cloud is also presented (blue).
- The coordinate is the object coordinate, object segment is viewed in the camera coordinate
Train a network
- Training data is created synthetically using 3D object model and 6D poses.
- The 6D pose and class id of target object are in ycb_video_data_tfRecords/train_syn/
- The data synthesis pipeline takes the target 3D object model and creates a segment of the object in the desired 6D pose. Below is two examples of synthetic segment (red), two real segments (red) are also shown for comparison.
Log files and trained model is store in log