Austin McEver

Welcome to my page :) I have just completed my PhD with the Vision Research Lab at the University of California, Santa Barbara after 5 years of computer vision research. My dissertation focuses on enhancing object detection of invertebrate species in a challenging, partially labelled dataset of underwater video that I have made public. I am also widely interested in other machine learning (ML) problems, and I enjoyed enhancing Facebook's recommender system for video during my summer internship in 2021.


Interests: weak supervision, partial supervision, computer vision, recommender systems, machine learning, deep learning

Contact: mcever ⓐt ucsb.edu

LinkedIn: https://www.linkedin.com/in/austinmcever/

Education

Professional Experience

Publications

Research Projects

Marine Applied Research and Exploration (MARE) has collected hundreds of hours of video using their unmanned, underwater, remote-operated vehicles (ROVs). In order to better survey and understand life in California's costal waters, MARE has annotated each video with species and substrate labels. 

My implementation of a Context-Driven Detector enables detection, tracking, and counting of invertebrate species with partial labels, while simultaneously generating temporal labels for substrates present in the MARE's videos as demonstrated on the soon-to-be public Dataset for Underwater Invertebrate Analysis (DUSIA).

Context-Matched Mosaic Generation for DUSIA

It can be difficult to collect images for training computer vision models, not to mention the costs associated with collecting annotations suitable for training these object detectors, like on DUSIA where annotations are partial and limited. To aid in the challenges associated with training with less supervision, Context Matched Collages leverage explicit context labels to combine unused background examples with existing annotated data to synthesize additional training samples that ultimately improve object detection performance. By combining a set of our generated collage images with the original training set, we see improved performance using three different object detectors on DUSIA, ultimately achieving state of the art object detection performance on the dataset.

Habitat Recognition in Underwater Vehicle Video

As a first step toward creating the Context-Driven Detector, I first implemented a method using a convolutional neural network (CNN) capable of generating temporal labels for DUSIA. A ResNet-based classification models classifies the video frame by frame, and a median filter temporally smoothes the classification results.

Semantic Segmentation of Underwater Images of Sessile Organisms

The Marine Science Institute at UCSB regularly photograph the sea floor near the Channel Islands and annotate their data using point labels, indicating the species present in a small group of pixels. I adapted a weakly supervised segmentation network designed for natural images to work with their data. A first draft of this work is available on arxiv at https://arxiv.org/abs/2007.05615.

BisQue

BisQue is an open source platform for storing, visualizing, organizing, and analyzing images in the cloud, largely maintained by UCSB’s Vision Research Lab. My responsibilities include software engineering tasks such as integrating Keycloak user authentication, developing machine learning modules for public use, assisting with Docker issues, and advising undergraduate research students who participate in the project.

Deep Superpixel Features

Superpixels refer to an over segmentation used to group pixels into homogenous regions based on some criteria. They have numerous applications, but describing them computationally is not as straightforward or powerful as it could be. I am working to adapt new convolutional neural networks (CNNs) that generate superpixels to simultaneously generate deep features for those superpixels.

Skills / Frameworks

Graduate Courses