This is the repository of the dataset and video material for the paper TGF-Net: Sim2Real Transparent Object 6D Pose Estimation Based on Geometric Fusion.
Google drive link
The proposed dataset contains ten kinds of objects, all of which are common types of objects in households. In order to include as many types of objects as possible, the ten objects include 5 symmetrical objects and 5 asymmetrical objects.
Since the synthetic dataset is made without noise such as motion blur and camera distortion, it is purer than the real dataset. We select 400 images of each object for training, which is large enough to train an accurate 6D pose estimation network. The backgrounds of the 400 images in the train dataset are generated by extracting one frame per ten frames of a video. At the same time, we extract 5000 backgrounds from another completely different video, and use these 5000 backgrounds to randomly generate a test dataset. We generate 2800 images for each object as a test dataset, so the entire dataset contains 32000 images, which we name Trans6D-32K. The test dataset and the train dataset do not have exactly the same background, and using different videos in the train dataset and the test dataset can make the background gap between them larger, so as to verify whether our method can resist the interference of the background by learning geometric features.
Download the 6D pose dataset Trans6D-32K
The structure of dataset folder should look like below, #01,#02,#03,#04,#05 are asymmetric objects, #06,#07,#08,#09,#10 are symmetry objects:
datasets/
The randomly generated object names are for easier data processing in dataset rendering, and they correspond to numbers #01 to #10 one by one.
├── Trans6D-32K
├──#01(budui1) # Object serial number and corresponding name
│ ├─test_2800 # Test dataset
│ │ ├─mask # mask
│ │ ├─pose # pose
│ │ └─rgb # RGB image
│ ├─train
│ │ ├─edge # edge information
│ │ ├─mask # mask
│ │ ├─pose # pose
│ │ ├─rgb # RGB image
│ │ └─xyz_crop # XYZ map
│ ├─bbox3d_and_center.txt # 3D bbox and center point of object
│ ├─camera.txt # camera intrinsics
│ ├─fps_64.txt # The 64 keypoints on the object
│ ├─model.ply # Model of object
│ ├─test_2800.json # Test dataset information in coco form
│ ├─train_400.json # Train dataset information in coco form
├──#02(budui2)
├──#03(budui3)
├──#04(budui4)
├──#05(budui5)
├──#06(bigcup)
├──#07(mcup31)
├──#08(mmmmm1)
├──#09(mmmmm2)
├──#10(zhuixin)
├──readme.md
├──models_info.json # Contains information about the size of the object