Object instance segmentation using recurrent models
Year: 2017
Type: dissertation
Abstract: This work intends to expand the Polygon-RNN model for object segmentation, able to create a polygon outlining an object using a CNN and a RNN, by trying several modifications. Several models and approaches have been explored. Out of them, the one that has shown the best IOU has been a joint model combining CNNs and a RNN. While the accuracy when running the model in default mode prediction only reaches 55.14% (which in fact is quite high), when running in beam search mode the accuracy has increased up to 59,94%. Once the final model was chosen, it was implemented in the LabelMe annotation tool, a web tool that can be used for annotators to segment instances of objects. This way the annotation process is made in a semi-automatic manner, making the work much faster and easier for the annotator. The annotator has to select a bounding box around the object and just slightly correct the resultant prediction to obtain the final annotation. The qualitative results have proven to be very precise.
Authors: Morales Vega, Juán Carlos
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: University of Toronto
OpenAlex ID: W2766709200