A Comparison of Deep Saliency Map Generators on Multispectral Data in Object Detection

Jens Bayer, David Münch, Michael ArensarXiv 2021 · arXiv

Abstract

Explainable AI (XAI) is crucial for high-stakes applications like autonomous driving, yet most research focuses on visible-light classification. This study evaluates three saliency map generators—Grad-CAM, RISE, and SIDU—on multispectral object detection tasks. By analyzing performance across RGB and various infrared spectra, the research demonstrates that multi-modal training and infrared data lead to more precise, interpretable model explanations.

Introduction

Deep learning models are often criticized as "black boxes," making it difficult to understand the reasoning behind their predictions. In multispectral object detection, where models process both visible (RGB) and infrared (IR) data, transparency is vital for safety. This paper addresses the gap in XAI research by systematically comparing saliency methods in complex, multi-modal environments.

Methodology

The study evaluates three distinct saliency map generators adapted for the EfficientDet-D0 object detector:

  1. Grad-CAM (Gradient-weighted Class Activation Mapping): A gradient-based method that is computationally efficient but often less precise.
  2. RISE (Randomized Input Sampling for Explanation): A perturbation-based method that randomly masks input regions. It is highly precise but computationally expensive.
  3. SIDU (Similarity Difference and Uniqueness): A newer approach that uses internal feature maps as masks to calculate regional contributions.
Comparison of saliency map generators showing varying levels of focus and noise across different methods.

Datasets and Multispectral Analysis

The researchers utilized the Multispectral Object Detection Dataset, featuring synchronized scenes in RGB, Near-Infrared (NIR), Mid-wave Infrared (MIR), and Far-Infrared (FIR/Thermal). A key finding was that saliency maps for infrared spectra tend to be more focused on the actual object than RGB maps, likely due to the high contrast of heat signatures.

Example of the Near-Infrared (NIR) spectrum from the Multispectral Object Detection Dataset, highlighting the high contrast of objects compared to visible light.

Key Insights

  • Multi-modal Synergy: Training on combined RGB and MIR data significantly improves explanation quality, resulting in tighter maps with less background noise compared to single-spectrum models.
  • Computational Trade-offs: There is a massive disparity in speed. Grad-CAM offers near real-time performance (0.1s), whereas RISE requires 32s per image, making Grad-CAM the only viable choice for real-time applications like moving vehicles.

Results

Performance was quantified using Deletion and Insertion metrics. SIDU achieved the best (lowest) deletion score of 0.10 ± 0.03 when used with NIR data. Furthermore, the combination of SIDU and RGB+MIR data yielded the highest insertion score (0.67 ± 0.21), indicating it most accurately identified the pixels that drive model confidence.

Conclusion

While SIDU and RISE provide more mathematically robust and precise explanations, their high computational cost remains a barrier for real-time deployment. Future work should focus on optimizing these perturbation-based methods and extending evaluations to spatio-temporal (video) and 3D volumetric data.

Summary updated

Abstract

Deep neural networks, especially convolutional deep neural networks, are state-of-the-art methods to classify, segment or even generate images, movies, or sounds. However, these methods lack of a good semantic understanding of what happens internally. The question, why a COVID-19 detector has classified a stack of lung-ct images as positive, is sometimes more interesting than the overall specificity and sensitivity. Especially when human domain expert knowledge disagrees with the given output. This way, human domain experts could also be advised to reconsider their choice, regarding the information pointed out by the system. In addition, the deep learning model can be controlled, and a present dataset bias can be found. Currently, most explainable AI methods in the computer vision domain are purely used on image classification, where the images are ordinary images in the visible spectrum. As a result, there is no comparison on how the methods behave with multimodal image data, as well as most methods have not been investigated on how they behave when used for object detection. This work tries to close the gaps. Firstly, investigating three saliency map generator methods on how their maps differ across the different spectra. This is achieved via accurate and systematic training. Secondly, we examine how they behave when used for object detection. As a practical problem, we chose object detection in the infrared and visual spectrum for autonomous driving. The dataset used in this work is the Multispectral Object Detection Dataset, where each scene is available in the FIR, MIR and NIR as well as visual spectrum. The results show that there are differences between the infrared and visual activation maps. Further, an advanced training with both, the infrared and visual data not only improves the network's output, it also leads to more focused spots in the saliency maps.
BibTeX
@misc{bayer2021comparison,
  author = {Jens Bayer and David Münch and Michael Arens},
  title = {{A Comparison of Deep Saliency Map Generators on Multispectral Data in Object Detection}},
  year = {2021},
  eprint = {2108.11767},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.11767},
}
Figure