Blog

Active Learning and the 2023 SPIE Conference

Expedition Technology · May 15, 2023

SPIE Defense + Commercial Sensing conference banner.

by Aimee Moses, Kelsey O’Haire, and Chris Bogart

At Expedition Technology (EXP), we ride the cutting edge of research to develop the most impactful solutions to our customers’ most difficult challenges. This month, three members of the Intelligent Classification in 3D (IC3D) team (Aimee Moses, Kelsey O’Haire, and Chris Bogart) attended the SPIE Defense and Commercial Sensing conference in Orlando, Florida to present our recent advances and learn about the latest developments by others in the research community.

Three colleagues standing in front of a large blue banner reading SPIE Defense and Commercial Sensing, wearing conference badges.

One of IC3D’s primary objectives is to design and build a solution for identifying and classifying objects in Light Detection and Ranging (LiDAR) and Synthetic Aperture Radar (SAR) 3D point clouds. Training effective object detectors requires a large amount of labeled point cloud data, which can be hard to come by. To address this challenge, we are applying a strategy called active learning, outlined in the figure below, which provides a framework for efficiently labeling data by identifying the most valuable data samples to label first.

Flow diagram of an active learning loop. A large pool of unlabelled samples feeds a selection step that picks a small subset to label; the minimally labelled dataset trains a model, which then runs inference back over the unlabelled pool to choose what to label next, converging on an optimally labelled dataset.

Aimee Moses presented the results of applying a variety of active learning approaches to the IC3D task in a talk titled, “Diversity-Based Active Learning: Creating a Representative Object Detection Dataset in 3D Point Clouds.” The accompanying paper the team wrote is available in the conference proceedings. We anticipate our findings will be applicable to a wide range of current and future projects at EXP and should be applicable to a broad set of machine learning projects with a lot of data, but very little of it labeled.

A presenter at a lectern beside a projected conference slide titled “Uncertainty Methods”, listing classification confidence and Shannon entropy as ways to find samples near a decision boundary, with the limitation that selected samples may be redundant.

In addition to Aimee’s presentation, we had the opportunity to listen to talks on topics ranging from unsupervised representation learning to underwater fish detection to automated classification of misinformation and more. By our calculations, between the 3 of us from EXP who attended the conference, we heard over 150 presentations! We are excited to share the new techniques and applications we learned about with the rest of our EXP colleagues and leverage them for the benefit of our customers.

Two sandhill cranes standing on a sunlit paved walkway outside a building, one preening its feathers.

Feathered visitors

Palm trees silhouetted against an orange sunset, seen through a sheer hotel curtain.

Florida Sunset from the Hotel

A landscaped indoor spring with clear shallow water and tropical planting; small alligators rest on the rocks on the far bank and a turtle sits on the sand in the foreground.

Alligators and turtles at Gator Springs