This page contains utility functions to use the DSEC-Detection datast. It is based off of the original DSEC dataset, but has added object detections in the left camera view. A preview of these labels, and download links can be found here. When using this dataset please cite the following two papers:

Daniel Gehrig, Davide Scaramuzza, "Low Latency Automotive Vision with Event Cameras", Nature, 2024 Open Access PDF.

@Article{Gehrig24nature,
  author    = {Gehrig, Daniel and Scaramuzza, Davide},
  title     = {Low Latency Automotive Vision with Event Cameras},
  booktitle = {Nature},
  year      = {2024}
}

and the original DSEC, on which this extension is based. Mathias Gehrig, Willem Aarents, Daniel Gehrig and Davide Scaramuzza, "DSEC: A Stereo Event Camera Dataset for Driving Scenarios", RA-L, 2021

@InProceedings{Gehrig21ral,
  author  = {Mathias Gehrig and Willem Aarents and Daniel Gehrig and Davide Scaramuzza},
  title   = {{DSEC}: A Stereo Event Camera Dataset for Driving Scenarios},
  journal = {{IEEE} Robotics and Automation Letters},
  year    = {2021},
  doi     = {10.1109/LRA.2021.3068942}
}

To set up the DSEC-Detection dataset, you need to

1. download the original dataset, let us denote the path to this dataset with $DSEC_ROOT
2. download the DSEC-Detection including the new sequences into the existing DSEC dataset
3. remap the images into the event view, or alternatively download the remapped images
4. test alignment

To install run

git clone [email protected]:uzh-rpg/dsec-det.git
cd dsec-det/

mamba create -n dsec-det python=3.7
mamba activate dsec-det
pip install -e .

mamba install -y -c conda-forge h5py blosc-hdf5-plugin opencv tqdm imageio pyyaml numba seaborn

Run the following commands to download the original DSEC dataset. The individual files can be found on the official DSEC project webpage to download the dataset to $DSEC_ROOT.

DSEC_ROOT=/path/to/dsec/
bash scripts/download_dsec.sh $DSEC_ROOT # $DSEC_ROOT is the destination path

Run the following command to download the extra data (events, images, etc.) and object detection labels to $DSEC_ROOT

bash scripts/download_dsec_extra.sh $DSEC_ROOT

Since images of DSEC are given in the left rectified image view, and labels are in the distorted event view, we need to remap the images to the event view. You can simply download them with the following commands:

bash scripts/download_remapped_images.sh $DSEC_ROOT

This will generate a new subfolder in $DSEC_ROOT/$split/$sequence/images/left/distorted, where the distorted images are stored.

You can now test alignment by running the following visualization script:

python scripts/visualize_example.py --dsec $DSEC_ROOT --split test

and this will load random samples from the dataset by generating a DSECDet dataset class. Feel free to use that class for your deep learning applications. The syntax looks like this:

import cv2 
from pathlib import Path 

from dsec_det.dataset import DSECDet
from dsec_det.io import yaml_file_to_dict


split_config = yaml_file_to_dict(Path("./config/train_val_test_split.yaml"))

dataset = DSECDet(root=Path("path/to/dsec_root"),
                  split="test",              # can be test/train/val
                  sync="back",               # load 50 ms of event after ('back'), or before  ('front') the image
                  split_config=split_config, # which sequences go into train/val/test. See yaml file for details.
                  debug=True)                # generate debug output, available in output['debug']
        
index = 5574
output = dataset[index]

cv2.imshow("Debug", output['debug'])
cv2.waitKey(0)

The output should look something like this:

The new sequences are summarized below and follow the same naming convention as DSEC

.
├── test
│   └── thun_02_a
└── train
    ├── zurich_city_16
    ├── zurich_city_17
    ├── zurich_city_18
    ├── zurich_city_19
    ├── zurich_city_20
    └── zurich_city_21

For all sequences, including the new ones, we provide object labels in the object_detections subfolder

sequence_name/
├── object_detections
│    └── left
│        └── tracks.npy
├─... 

Each tracks.npy file contains all the object detection, with associated track id for that sequence. It is stored as a numpy array, following the format by Prophesee. The keys are:

t:                (uint64)  timestamp of the detection in microseconds.
x:                (float64) x-coordinate of the top-left corner of the bounding box
y:                (float64) y-coordinate of the top-left corner of the bounding box
h:                (float64) height of the bounding box
w:                (float64) width of the bounding box
class_id:         (uint8)   Class of the object in the bounding box. 
                            The classes are ('pedestrian', 'rider', 'car', 'bus', 'truck', 'bicycle', 'motorcycle', 'train')
class_confidence: (float64) Confidence of the detection. Can usually be ignored.
track_id:         (uint64)  ID of the track. Bounding boxes with the same ID belong to one track.