Datasets | eCo-DRIVERS

CVC-09 FIR Sequence. Pedestrian Dataset

The sequence is composed by two sets of images, named as the day and night sets, which refers to the moment of the day they were acquired. The first set contains 5990 frames and the second 5081, divided in training and testing sets each sequence.

Here are some examples of the dataset:

References to this pedestrian dataset should be made to the following article:

Adapting Pedestrian Detection from Synthetic to Far Infrared Images:
- Yainuvis Socarras, Sebastian Ramos, David Vazquez, Antonio Lopez, Theo Gevers. Adapting Pedestrian Detection from Synthetic to Far Infrared Images, In ICCV – Workshop on Visual Domain Adaptation and Dataset Bias. Sydney, Australy, 2013

Download: CVC-09 FIR Sequence (1.87 Gb)

Please, read the terms of use before downloading the dataset: disclaimer

CVC-08 On-board Sequence. Pedestrian Dataset

The sequence was acquired on-board under normal urban driving conditions. The images are monochrome and of 480X960 pixels. We used a 4mm focal length lens, so providing a wide field of view. We drove during 30 minutes approximately, giving rise to a sequence of around 60,000 frames. Then, using steps of 10 frames we annotated all the pedestrians. This turns out in 7,900 annotated pedestrians, 5,400 reasonable and non occluded. We have divided the video sequence into three sequential parts, the first one for training, the last one for testing, in the middle we have leaved a gap for avoiding testing and training with the same persons. Overall we train with 3,600 reasonable pedestrians, and test on 1,300 reasonable ones.

Here are some examples of the dataset:

References to this pedestrian dataset should be made to the following article:

Spatiotemporal Stacked Sequential Learning for Pedestrian Detection:
- A. Gonzalez, S. Ramos, D. Vazquez, A. M. Lopez, J. Amores. Spatiotemporal Stacked Sequential Learning for Pedestrian Detection, In 7th Iberian Conference on Pattern Recognition and Image Analysis (IbPRIA), 2015

Download: CVC08 On-board Sequence Pedestrian Dataset (1.96 GB)

Please, read the terms of use before downloading the dataset: disclaimer

CVC-07 DPM Virtual-World Pedestrian Dataset

This dataset contains 2534 pedestrian images and 2000 background images. The pedestrian images have frontal view and left view, which are annotated as ‘M’ and ‘L’. You may flip the pedestrians to get right view examples. Part annotations are also provided.

Here are some examples of the dataset:

References to this pedestrian datset should be made to at least one of the following articles:

Advantages of highly aligned (views and parts) training data:
- J. Xu, D. Vazquez, A. M. Lopez, J. Marin, D. Ponsa, Learning a Part-based Pedestrian Detector in Virtual World. In IEEE Transactions on Intelligent Transportation Systems (T-ITS), 2014.
Domain adaptation of DPM (virtual to real included):
- J. Xu, S. Ramos, D. Vazquez, & A. M. Lopez, Domain Adaptation of Deformable Part-Based Models. In IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 2014.
Domain adaptation of with active learning (especial focus on HOG and LBP):
- D. Vázquez, J. Marín, A.M. López, D. Ponsa and D. Gerónimo, Virtual and Real World Adaptation for Pedestrian Detection. In IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 2014.

Download: CVC07 DPM Virtual-World Dataset (1.3GB)

Please, read the terms of use before downloading the dataset: disclaimer
More info here

CVC-06 Partially Occluded Virtual-World Pedestrian Dataset

This is a partially occluded pedestrian dataset extracted from our virtual world. The dataset consists of two sets of 2400 and 2414 croped pedestrians (with their corresponding horizontal mirrors).

Here are some examples of the dataset:

References to this pedestrian dataset should be made to the following article:

J. Marín, D. Vázquez, A.M. López, J. Amores and L.I. Kuncheva, Occlusion handling via random subspace classifiers for human detection. In IEEE Transactions on Systems, Man, and Cybernetics (Part B) (ITSMC-B), 2013.

Download: CVC06-Partially Occluded Pedestrian Dataset (137MB)
Please, read the terms of use before downloading the dataset: disclaimer

CVC-05 Partially Occluded Pedestrian Dataset

This is a partially occluded pedestrian dataset. The dataset consists of 593 positive frames with annotated pedestrians (with their corresponding horizontal mirrors).

Here are some examples of the dataset:

References to this pedestrian datasetshould be made to the following article:

J. Marín, D. Vázquez, A.M. López, J. Amores and L.I. Kuncheva, Occlusion handling via random subspace classifiers for human detection. In IEEE Transactions on Systems, Man, and Cybernetics (Part B) (ITSMC-B), 2013.

Download: CVC05-Partially Occluded Pedestrian Dataset (274MB)
Please, read the terms of use before downloading the dataset: disclaimer

CVC-04 Virtual-World Pedestrian Dataset 2

This is the virtual pedestrian dataset, generated using Half-Life 2 graphics engine. The dataset consists of 1208 virtual pedestrians (with their corresponding horizontal mirrors) and 6828 pedestrian-free background images to extract negatives for training.

Here are some examples of the dataset:

References to this pedestrian datasetshould be made to the following article:

D. Vázquez, J. Marín, A.M. López, D. Ponsa and D. Gerónimo, Virtual and Real World Adaptation for Pedestrian Detection. In IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 2014.

Download: CVC04-Virtual-Pedestrian Dataset 2 (2,86GB)
Please, read the terms of use before downloading the dataset: disclaimer

CVC-03 Virtual-World Pedestrian Dataset

This is the virtual pedestrian dataset, generated using Half-Life 2 graphics engine. The dataset consists of 1678 virtual pedestrians (with their corresponding horizontal mirrors) and 2048 pedestrian-free background images to extract negatives for training.

Here are some examples of the dataset:

References to this pedestrian datasetshould be made to the following article:

J. Marín, D. Vázquez, D. Gerónimo and A.M. López, Learning appearance in virtual scenarios for pedestrian detection. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), San Francisco, CA, USA, 2010.

Download: CVC03-Virtual-Pedestrian Dataset (915MB)
Please, read the terms of use before downloading the dataset: disclaimer

CVC-02 Pedestrian Dataset

CVC-02 consists of three subsets, each one focused on a different task of pedestrian detection: candidate generation, classification and system evaluation. The imagery has been recorded in urban scenarios around Barcelona (Spain), using a Bumblebee color stereo camera with resolution 640×480 pixels and 6mm focal length. The annotated pedestrians are in the range from 0 to 50 m from the camera, which corresponds to a smallest pedestrian of 12×24 pixels. The main features of each subset are the following:

CVC-02-CG (Candidate Generation): 100 frames in color, depth and 3D points information.
CVC-02-Classification: training (1016 cropped positives and mirrors, 7650 cropped negatives and pedestrian-free images), testing set (window-based, with 570 cropped positives and mirrors, 7500 cropped negatives and pedestrian-free images) and testing set (image-based, with 250 frames containing 587 annotated pedestrians).
CVC-02-System: 15 sequences of 4364 frames, 7983 pedestrian annotations.

All the images are provided in lossless PNG format, both in color and depth versions, and in their original size and 64×128 pixels rescaled. Regarding the annotations, we label them as obligatory or optional (very young children, significantly occluded or partially out of the image).

Here are some samples of the dataset:

References to this pedestrian datasetshould be made to the following article:

D. Gerónimo, A.D. Sappa, D. Ponsa and A.M. López. 2D-3D Based on-board pedestrian detection system. In Computer Vision and Image Understanding (Special Issue on Intelligent Vision Systems) (CVIU) Vol. 114, Num. 5, pp. 583-595. May 2010.

Download: CVC-02-CandidateGeneration (269MB)
Download: CVC-02-Classification (1.5GB)
Download: CVC-02-System Sequence 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 (around 1GB each)
Since the complete dataset size is around 12GB we have partitioned it into smaller files.
Please, read the terms of use before downloading the dataset: disclaimer

CVC-01 Pedestrian Dataset

This pedestrian datasetwas recorded in the surreoundings of Barcelona and annotated by the CVC ADAS group in order to evaluate the pedestrian detection algorithms developed. It was originally named CVC-CER-01. Its main features are the following:

1000 manually annotated pedestrians (with the corresponding 1000 mirror images)
6175 random non-pedestrians windows from the road area (i.e., no sky or easy samples)
Original scale PNG files

Here are some scale-normalized pedestrian examples of the dataset:

References to this pedestrian datasetshould be made to the following article:

D. Gerónimo, A.D. Sappa, A. López, D. Ponsa. Adaptive image sampling and windows classification for on-board pedestrian detection. In Proceedings of the International Conference on Computer Vision Systems (ICCVS), Bielefeld, Germany, 2007.

Download: CVC-01 Classification Dataset (83MB)
Please, read the terms of use before downloading the dataset: disclaimer

Road Image Dataset

This database addresses the need for experimental data to quantitatively evaluate emerging road detection and photometric invariant algorithms. These images were captured using an onboard camera mounted on the windshield of a moving vehicle. The main challenge is dealing with lighting variations and shadows. The database consists (right now) of two sequences acquired on the same scenario at different daytime and under different weather conditions. The first sequence was acquired at noon of a sunny day. The second was acquired in the morning after raining. The camera is a Bumblebee (stereo pairs will be available soon) working with fixed parameters. Road groundtruth has been manually generated. Here are some examples of the database:

Day sequence	Rain sequence

References to this database should be made to the following article:

Road Detection Based on Illuminant Invariance. José M. Álvarez and Antonio M. López. In IEEE Transactions on Intelligent Transportation Systems, 2011.

Download: Day Sequence and Rain Sequence