Human Depth Dataset

This repository provides some tools for analyzing and extracting subsets the KITTI dataset with emphasis on pedestrians and depth data.

This also provides a convenient dataloader for the RGB-D people dataset.

Usage

KITTI

Preparing the data

• Download the depth prediction ground truth data from KITTI's website.
• Download the corresponding raw data using depth_val_raw_data_downloader.sh (Only downloads the raw data corresponding to the depth validation set to save space).
• Move the contents of the validation set from ground-truth depth to data/kitti/val/gt.
• Restructure the raw data under data/kitti/val/raw so that the final structure looks like this:

data/
--- val/
--- --- gt/
--- --- --- 2011_09_26_drive_0002_sync/
--- --- --- 2011_09_26_drive_0005_sync/
--- --- ...
--- --- raw/
--- --- --- 2011_09_26_drive_0002_sync/
--- --- --- 2011_09_26_drive_0005_sync/
--- --- ...

Extracting statistics

You can use extract_human_stats.py to run human detection on every frame of 'camera 2' from the raw data and save the results.

Afterwards you can use the anayze_stats notebook to analyze the results, and create a list of images with a minimum number of people in them (scenes_with_min_2_people.txt for example).

Dataloader

Inside dataset.py a pytorch dataset is defined that given a list like the one produced above, will generate RGB and ground-truth depth pairs for evaluation.

RGB-D People

Doenload the dataset and extract in data/rgbd/

Dataloader

Use RGBDPeopleDataset in dataset.py.