Del Ser, Javier (TECNALIA), Martinez Seras, Aitor (TECNALIA), Laña, Ibai (TECNALIA), Bilbao, Miren Nekane (University of the Basque Country), Fafoutellis, Panagiotis (National Technical University of Athens), Vlahogianni, Eleni I. (School of Civil Engineering, National Technical, University of A)

On the Connection between Neural Activations and Uncertainty in Object Detection Transformers Using Topological Data Analysis

Scheduled for presentation during the Regular Session "Sensing, Vision, and Perception I" (WeAT2), Wednesday, September 25, 2024, 11:10−11:30, Salon 5

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

Over the years, deep neural networks have achieved unrivaled levels of predictive performance in detecting and identifying objects from visual data, elevating them as a core technology for vehicular perception and automated driving. Recently, the research interest has drifted from performance-driven advances towards the improvement of the reliability and robustness of neural object detectors when operating in open-world learning scenarios. In this context, much attention has been paid especially to their capability to detect out-of-distribution objects from their input image. This work aligns with this rising concern by proposing a methodology to compute, represent and examine the topology of neural activations triggered by objects detected in an image by an object detection model. Our methodology allows examining the geometry of such activations, the semantics of objects sharing similar activation patterns, and the estimated confidence of the model when detecting such objects. The overall aim of the methodology is to identify activation vectors that are indicative of reliable detection, and anomalous activation patterns that may signify out-of-distribution objects. Our experiments with a pretrained object detection Transformer and vehicular image data expose a close link between the cohesiveness of neural activation patterns for known object categories, the confidence of the model in the prediction of objects belonging to such known categories, and the type of object itself (the semantics). These findings confirm that neural activations can be used to detect novel or unforeseen objects at the model's input image.