A new object proposal generation method for object detection in RGB-D data

This paper proposes a modified selective search method that generates object proposals on RGB-D data in indoor scenes. The proposed method first applies color flattening to generate monotonous color variations in RGB image data. Then, from the color-flattened image and depth map data, cost function-based segment grouping and depth segmentation are applied to produce desirable segmentation results. Segment grouping using cost function on image data computes dissimilarities in color, texture, and size between two adjacent regions with pre-learned weights. Depth segmentation uses the height difference of grid cells in the binned depth grid map. The final set of object proposal regions extracted from the RGB image and depth map data is organized by considering the overlapping between two data modalities. Finally, the extracted set of object proposal regions is fed into AlexNet or VGG-16, both of which are widely used for object classification, to evaluate our method on object detection and classification tasks. The proposed segment-based method can precisely detect meaningful object regions using a smaller number of proposals than other methods. Further, its detection and classification performance are better than those of previous methods.