A repo to summarize datasets used for object pose estimation and rendering methods used to generate synthetic training data.
In the following tables, 3D CAD model is noted as model and 2D pictured object is noted as object.
Papers resuming some related datesets can be found here and here
This table lists the datasets commonly known as BOP: Benchmark 6D Object Pose Estimation, which provide accurate 3D object models and precise 2D-3D alignment.
You can download all the BOP datasets here and use the toolkit provided by the authors.
You can use our code ply2obj.py
to convert original .ply files to .obj files,
and run create_annotation.py
to create a single annotation file for all the scenes in a dataset.
Datasets format can be found here, we use instance id in our annotation to indicate different instances pictured in the same image.
Dataset | Sample image | Annotation | Statistics | Reference |
---|---|---|---|---|
HomebrewedDB | 6D pose + Depth + BoundingBox | 33 models in 13 videos with 17,420 frames | Preprint 2019 | |
YCB-Video | 6D Pose + Depth + Mask | 21 models in 92 videos with 133,827 frames | RSS 2018 | |
T-LESS | 6D Pose + Depth | 30 models in 20 videos with ~49K frames | WACV 2017 | |
Doumanoglou | 6D Pose + Depth | 2 models in 3 videos with 183 frames | CVPR 2016 | |
Tejani | 6D Pose + Depth | 6 models in 6 videos with 2,067 frames | ECCV 2014 | |
Occluded-LINEMOD | 6D Pose + Depth | 8 models in 1,214 frames with 8,992 objects | ECCV 2014 | |
LINEMOD | 6D pose + Depth for one object | 15 models in 15 videos with 18,273 frames | ACCV 2012 |
In this table, Pix3D and ScanNet provide precise 2D-3D alignment while others only provide a coarse alignment.
PASCAL3D+ is the de facto benchmark used for viewpoint estimation.
ScanNet is usually used to evaluate scene reconstruction and segmentation.
Dataset | Sample image | Annotation | Statistics | Reference |
---|---|---|---|---|
ApolloCar3D | 6D Pose + Mask | 34 car models with 60K+ objects in 5,277 images | CVPR 2019 | |
Pix3D | 6D Pose + Mask | 9 categories containing 395 models in 10,069 images | CVPR 2018 | |
ScanNet | 6D Pose + Segmentation + Depth | 2.5M RGB-D frames in 1,515 scenes | CVPR 2017 | |
ObjectNet3D | Euler Angles + BoundingBox | 100 categories with 201,888 objects in 90,127 images | ECCV 2016 | |
PASCAL3D+ | Euler Angles + BoundingBox | 12 categories with 36,292 objects in 30,889 images | WACV 2014 | |
KITTI | 3D BoundingBox | 80,256 objects in 14,999 images | CVPR 2012 |
In order to testify the network generalization ability (tested on images containing unseen 3D models from the training set), the following dataset could be used to generate synthetic training data.
Notice that ABC contains generic and arbitrary industrial CAD models while ShapeNetCore and ModelNet contain common category objects such as cars and chairs.
Dataset | Number of categories | Number of models | Reference |
---|---|---|---|
ABC | - | 1 million | CVPR 2019 |
ShapeNetCore | 55 | ~51,300 | ArXiv 2015 |
ModelNet-40 | 40 | 26,960 | CVPR 2015 |
Neural 3D Mesh Renderer: Kato el al. CVPR 2018
RenderNet: Thu et al. NIPS 2018
In this repo, we provide python code to generate rendering images from 3D models using blender as a python module that is easy to install and generate photo-realistic images : )
TODO: scripts about how to use it.
Other works using blender can be found here that generates one model at a time.
PyBullet: a very popular one in the Robotics community.
Glumpy: does not support headless rendering (failed on ssh mode)
UnrealCV: extension of Unreal Engine 4, helps interact with virtual world and communicate with external program.
SyntheticComputerVision: resuming a lot of techniques used to generate synthetic image
Attention: 3D models should be aligned in the same way through meshlab to ensure the consistent orientation while wandering across the different datasets.