Datasets

BEHAVE Action Dataset full annotation

BEHAVE Interactions Test Case Scenarios was presented in 2009 in the paper "The BEHAVE video dataset: ground truthed video for multi-person behavior classification" by S. Blunsden and R.B. Fisher from the University of Edinburgh. It consists in a 1 hour long surveillance video with actors performing different collective actions. The database was aimed as a benchmarking groundtruth for people interaction detection/classification algorithms. During my research, I found BEHAVE dataset very convenient also for forensic video analysis given that it contains different people reappearing and performing different actions through a relatively long video.

One of the problems of this dataset is that the annotations correspond to a small part of the video, and are only focused on the interaction itself. Here I provide the persons annotations in the whole video together (5 re-appearing actors plus other not re-appearing not-actors) with the individual actions (walking , fighting, running and standing).

The annotations are provided in a single txt file and the format is as follows:

         numObjects

         idObj type

         numFramesPositions

         frame xc yc w h

         frame xc yc w h

         frame xc yc w h

         frame xc yc w h

         ...

         numFramesActions

         frameStart frameEnd action

         frameStart frameEnd action

         frameStart frameEnd action

         frameStart frameEnd action

         ...

References to these annotations should be made to the following article:

D. Gerónimo and H. Kjellström. Unsupervised Unsupervised surveillance video retrieval based on human action and appearance. In Proc. Int. Conf. on Pattern Recognition, pp. 4630-4635, Stockholm, Sweden, 2014.

Download: Annotations and video.

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 640x480 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 12x24 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 64x128 pixels rescaled. Regarding the annotations, we label them as obligatory or optional (very young children, significantly occluded or partially out of the image).

References to this pedestrian database should 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), Vol. 114, Num. 5, pp. 583-595, 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)
Please, read the terms of use before downloading the database: disclaimer.

CVC-01 Pedestrian Dataset

This pedestrian database was 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

References to this pedestrian database should 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, Bielefeld, Germany, 2007.

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