The new database addresses the issues of variability of action-related spatio-temporal motion features over multiple actions, actors, objects, clothes, action contexts, scales, surroundings and views, many of which have not been addressed explicitly by the proposed action models neither were they addressed properly by the existing databases and applied testing methodologies.

Full-body action database contains several sequences of actors performing actions with the whole body in view. Actions were chosen to range from full-body actions (e.g. sitting on a table) to part-ofbody actions (e.g. picking up a cup with a hand). The following action sequences are recorded:

• moving a box: a person picking up a box from the floor, putting it on the table, picking it up again and releasing it on the floor
• making a phone call: a person standing, approaching the phone, making a phone call, and hanging-up
• drinking from a cup: a person sitting, picking up a cup, drinking, and returning the cup
• sitting on a chair: a person standing next to a table, sitting down, and standing up
• sitting on a table: a person standing next to a table, sitting on the table, and standing up again

The action sequences were recorded in three different locations within the Computer Vision Laboratory at the University of Ljubljana. Besides changing locations there is additional variability between shots:

• the illumination varied between shots with varying ratio of daylight and fluorescent illumination;
• most actors changed top clothes from shot to shot;
• the camera changed location from shot to shot;
• the scale varies slightly between shots;
• cups and box were different color in several shots;
• the actors assumed natural position next to the table.

For each action sequence at least 3 complete performances by 7 people were recorded in 3 different locations, resulting in at least 63 repetitions of each individual action. Another person was recorded in a single location only. A detailed description of the database is given in the deliverable DR.5.3.

The video files are named 'P - L - A.avi' where

• P is the person ID (001...008),
• L is the index of location (1...3) and
• A is action sequence descriptor (box, chair, cup, phone, table).

The contents are recorded in PAL DV quality, i.e. 720×576 resolution, 25 frames per second where each frame contains 2 interlaced half-frames. The size of video sequences varies from 582 frames (23.28 seconds) to 1324 frames (52.96 seconds). The file format is AVI, the video format FourCC code is 'dvsd'. The total size of the database is 12.8 gigabytes.

For download instructions, please send an e-mail to Miha Peternel. However, since the total size of the database is 12.8 gigabytes, the download can take a long time. Alternativelly, we can send you the database on three DVDs by mail on your request.

Print this page