Object class segmentation of RGB-D video using recurrent convolutional neural networks

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Object class segmentation is crucial for computer vision, assigning labels to image pixels.
• Deep Convolutional Neural Networks (DNNs) excel at local spatial correlations but struggle with long-range dependencies due to fixed filter sizes.
• Recurrent Neural Networks (RNNs) can model long-range dependencies through iterative activity propagation, beneficial for video analysis.

Purpose of the Study:

• To present a novel RNN architecture for object class segmentation.
• To investigate various training methodologies for this RNN architecture.
• To evaluate the performance of the proposed models on a standard dataset.

Main Methods:

• Development of a novel Recurrent Neural Network (RNN) architecture tailored for object class segmentation.
• Exploration and comparison of different training strategies for the RNN models.
• Evaluation using the NYU Depth v2 dataset, a benchmark for depth-based object class segmentation.

Main Results:

• The proposed RNN architecture demonstrates competitive performance on the NYU Depth v2 dataset.
• The models effectively address the limitations of DNNs in capturing long-range dependencies.
• Successful application of RNNs to pixel-level object class labeling tasks.

Conclusions:

• The novel RNN architecture is effective for object class segmentation, particularly in scenarios requiring the understanding of long-range spatial and temporal information.
• The findings suggest RNNs are a promising approach for enhancing segmentation tasks in computer vision.
• The research contributes competitive results to the field of object class segmentation using deep learning models.