MVAR Datasets

∗ MVAR stands for Multiple Viewpoints Action Recognition.

We would like to introduce two datasets for multiple viewpoints action recognition:

  1. Simulated dataset named MVAR-Unity3D Attack
  2. Realistic dataset named MVAR-GoPro Standing Exercise

In both datasets:

  • The videos have a resolution of 640×480 pixels.
  • The number of action classes is eight.

1. MVAR-Unity3D Attack

To record this dataset, a simulation on Unity3D software based on 3D models made by Studio New Punch, Niandrei Explosive, and RockVR was conducted. A camera was utilized and rotated around the roll axis (i.e., Y axis) by a step 5 deg from 0 to 90 deg to have 19 viewpoints. In this dataset, there are five characters with characteristic variables, such as height, head, body, and clothes. Each character performed each of the eight short action classes twice in each of eight scenarios of variations as listed in Table 1 and Table 2 bellow. In total, the dataset consists of 19*5*8*8*2=12,160 videos. There is 1,520 videos for each scenario. Especially, for each frame of every video, a random function was used to make a change within a specified range of movements (for all parts, e.g., hands, legs, and head) even though these videos belonged to the same action class (see movement variations in the second column of Figure 1 below. Note that these movement variations were applied to all frames in every video of all eight scenarios. The random function was also utilized to change the subject scale, position where the action was performed, action speed, lighting condition, location of partial occlusion, and different backgrounds.

Figure 1. Experimental setting and several examples of variations for simulated MVAR-Unity3D Attack dataset.

Table 1. Scenarios of variations and their video sizes in simulated MVAR-Unity3D Attack (C0 to C7) and realistic MVAR-GoPro Standing Exercise (CR) datasets.

Table 2. Action classes in MVAR-Unity3D Attack dataset.

2. MVAR-GoPro Standing Exercise

In this dataset, five GoPro cameras located at five viewpoints from 0 to 90 deg were utilized. The average difference between two neighbor viewpoints was 22.5 deg. There were six actors and eight action classes of standing exercises as listed in Table 3 below. Each action was performed twice by each actor.

Figure 2. Experimental setting for realistic MVAR-GoPro Standing Exercise dataset.

Table 3. Action classes in MVAR-GoPro Standing Exercise dataset.


MVAR-Unity3D Attack Dataset

MVAR-GoPro Standing Exercise Dataset