UnLoc dataset (Synthetic + Real)

This dataset contains the synthetic and real data used in the article " Virtual Training for a Real Application: Accurate Object-Robot Relative Localization Without Calibration " to train 3 convolutional neural networks (CNNs) in order to perform uncalibrated relative localization of a cuboid block with respect to a robot, and to evaluate them. 

It consists of a dataset composed of synthetic pictures for training the CNNs and a dataset of real pictures for evaluation. 

The "synthetic" dataset is composed 3 sub-datasets (each of them composed of thousands of synthetic pictures and corresponding groundtruth) for training : 

1. a dataset for coarse relative localization subtask : coarse_estimation_data (~14.6 GB when extracted)
2. a dataset for the tool localization subtask : tool_detection_data (~4.3 GB when extracted)
3. a dataset for the fine relative localization subtask : fine_estimation_data (~13.3 GB when extracted)

These sub-datasets are composed of raw data as well as post-treated data ready to be used for training CNNs, in CSV format and in Torch format (.t7). 

The "real" dataset is composed of real pictures with a precisely localized cuboid block for evaluation only : UnLoc_real (~2.5 GB when extracted). 

More information available on the project page : http://imagine.enpc.fr/~loingvi/unloc/