liuguoyou/odms

Object Depth via Motion and Segmentation Dataset

ODMS Dataset

ODMS is the first dataset for learning Object Depth via Motion and Segmentation. ODMS training data are configurable and extensible, with each training example consisting of a series of object segmentation masks, camera movement distances, and ground truth object depth. As a benchmark evaluation, we also provide four ODMS validation and test sets with 15,650 examples in multiple domains, including robotics and driving. In our paper, we use an ODMS-trained network to perform object depth estimation in real-time robot grasping experiments, demonstrating how ODMS is a viable tool for 3D perception from a single RGB camera.

(New) An object detection-based version of the ODMS benchmark is now available here!

Contact: Brent Griffin (griffb at umich dot edu)

Quick Introduction: https://youtu.be/c90Fg_whjpI

Using ODMS

Run ./demo/demo_datagen.py to generate random ODMS data to train your model.
Example training data configurations are provided in the ./config/ folder. Has the option to save a static dataset.
[native Python, has scipy dependency]

Run ./demo/demo_dataset_eval.py to evaluate your model on the ODMS validation and test sets.
Provides an example evaluation for the VOS-DE baseline. Results are saved in the ./results/ folder.
[native Python, VOS-DE baseline has skimage dependency]

Benchmark

Method	Robot	Driving	Normal	Perturb	All
DBox	11.5	24.8	11.8	20.3	17.1
ODNlr	13.1	31.7	8.6	17.9	17.8
BoxLS	17.6	33.3	13.7	36.6	25.3
VOS-DE	32.6	36.0	7.9	33.6	27.5

Is your technique missing although it's published and the code is public? Let us know and we'll add it.

Publication

Please cite our paper if you find it useful for your research.

@inproceedings{GrCoECCV20,
  author = {Griffin, Brent A. and Corso, Jason J.},
  booktitle={The European Conference on Computer Vision (ECCV)},
  title = {Learning Object Depth from Camera Motion and Video Object Segmentation},
  year = {2020}
}

ECCV 2020 Presentation: https://youtu.be/ZD4Y4oQbdks

Use

This code is available for non-commercial research purposes only.