Skip to main content
Utah's Foremost Platform for Undergraduate Research Presentation
2021 Abstracts

Faster RCNN on Pelican Detection

Presenters: Swornim Chhetri, College of Computer Sciences, Computer Science
Authors: Swornim Chhetri, Yongtai Li, Daniel Salinas Duron
Faculty Advisor: Daniel Salinas Duron, College of Computer Sciences, Computer Science
Institution: Westminster College

Population tracking of wildlife can be done non-invasively using remote cameras. Over the breeding season, the volume of images produced is more than can be labeled efficiently by humans. Convolutional neural nets are the state-of-the-art solution for the problem of object detection in images. We trained a neural network based on the Faster Region-Based Convolutional Neural Network architecture (Faster-RCNN) to detect pelicans that inhabit the Great Salt Lake. We trained our network on images of pelicans and seagulls because both inhabit the island. Our training set consisted of three different classes: pelican, seagull, and background. During training, we applied transfer learning as well as redesigning the classification head. We built and trained our network using the Python package PyTorch. The PyTorch implementation of Faster R-CNN utilizes a pretrained ResNet 50 feature pyramid network for feature extraction by default. We also experimented with different pretrained feature extractors. Using a sub-sample of the images, we estimated with 95% confidence that the mean of the number of pelicans is within 3.85 to 4.87 per image. We use this statistic as a baseline for our network predictions. After initial training, we observed that the model had a high number of false positives, indicating underfitting. Experiments revealed that using a smaller feature extractor and using all module losses to train the network gave a noticeably better results. We plan to further refine the network architecture and use data augmentation and image preprocessing to improve our results further.