UNDERWATER BENCHMARK DATASET FOR TARGET DETECTION AGAINST COMPLEX BACKGROUND

The underwater benchmark dataset consists of 14 videos categorised into seven different classes representing complex challenges in the background modeling

  1. The “Blurred” sequences which aim at evaluating how smoothed and low contrasted images may affect the algorithms’ performance. This is particularly important for all those approaches that use textures in the background/foreground modeling process;
  2. The “Complex Background” sequences show cases where background is featured by complex textures and aim at assessing how approaches not using textures perform;
  3. The “Crowded” sequences aim at evaluating the algorithms’ performance with many occluding objects;
  4. The “Dynamic Background” and the “Luminosity Changes” videos are featured, respectively, by background objects movements (e.g. plant movements due to the marine currents) and transient and abrupt luminosity changes and aim at evaluating how good are the approaches in filtering out potential false positives;
  5. The “Camouflage Foreground Objects” sequences for evaluating the effects of camouflage.
  6. The “Hybrid” sequences, finally, show a combination of the above conditions.

The ground truth on the underwater video dataset was hand-labeled using the PERLa tool. For each video about 30 images were annotated and are provided as binary masks. In total more than 3500 objects were labeled. Please note that in the ground truths there are masks completely black, which are meant to assess performance when only background moving objects are present in the scene.

Blurred

Blurred

Video
Ground Truth

Complex Background

Complex

Video
Ground Truth

Crowded

Crowded

Video
Ground Truth

Dynamic Background

Dynamic

Video
Ground Truth

Hybrid

Hybrid

Video
Ground Truth

Luminosity Variations

Luminosity

Video
Ground Truth

Camouflage Foreground Objects

Camouflage

Video
Ground Truth



REFERENCE

Please cite as:

I. Kavasidis, S. Palazzo, R. Di Salvo, D. Giordano and C. Spampinato, An Innovative Web-Based Collaborative Platform for Video Annotation, Multimedia Tools and Applications, pp. 1 – 20, 2013.

NOTE

We have also an extensive dataset with about 40000 objects and 4000 trajectories for testing tracking algorithms on underwater domain. To obtain a copy of this dataset, please contact Dr. Concetto Spampinato (cspampin[at]dieei (dot) unict (dot) it)