A Benchmark Dataset and Saliency-Guided Stacked Autoencoders for Video-Based Salient Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Image-based salient object detection (SOD) is well-researched, but video-based SOD lacks large-scale, well-annotated datasets.
• Existing video SOD research is limited by the absence of unambiguous definitions and annotations for salient objects in dynamic scenes.

Purpose of the Study:

• To introduce the largest dataset to date for video-based salient object detection (SOD).
• To propose an unsupervised baseline method for video SOD using saliency-guided stacked autoencoders.
• To establish a challenging benchmark for advancing video SOD research.

Main Methods:

• A new dataset comprising 200 videos with manually annotated objects/regions on 7650 keyframes and eye-tracking data from 23 subjects was created.
• Salient objects were defined as those consistently appearing prominent throughout a video, annotated using combined manual masks and multi-subject eye-tracking data.
• An unsupervised approach employing saliency-guided stacked autoencoders was developed, extracting spatiotemporal cues at pixel, superpixel, and object levels.

Main Results:

• The proposed unsupervised method outperformed 30 out of 31 state-of-the-art models on the new dataset.
• The dataset proved highly challenging, demonstrating its potential to drive progress in video SOD.
• The method successfully inferred pixel saliency by encoding spatiotemporal cues from neighbors.

Conclusions:

• The developed dataset and unsupervised method represent a significant advancement for video salient object detection.
• The findings suggest that consistent object prominence is a key factor in defining video saliency.
• The proposed approach offers a strong baseline and benchmark for future research in unsupervised video SOD.