Literature Review: Superpixel Segmentation in Autonomous Vehicles
Here’s what I found while wrangling through current research on superpixel segmentation techniques. It reveals crucial findings that can be used to build a case in a later implementation of a more advanced AI model that overcomes limitations founds.
Enjoy the read…
In recent times, a sizeable amount of literature has focused on superpixel methods for visual explanations in solving multiple problems rare-class problems (Yu & Fan, 2021) and provide an end-to-end trainable algorithm (Wang et al., 2020). In an interesting article, (Hartley et al., 2021) succeeded with an approach called Super-pixel Weighted by Average Gradients (SWAG) and claimed it is more accurate, efficient and tunable drop in replacement method for GRAD-CAM. However, (Liu et al., 2020) uncovered superpixel methods’ arbitrary shape make them hard to use in convolutional neural networks (CNNs). Different studies have shown promise in overcoming associated problems of superpixel explanations; (Cai et al., 2021) found a way of reducing expensive annotation cost for sematic segmentation, while (Suzuki, 2020) proposed a segmentation method with quantitative and qualitative benefits when applied to CNNs. The studies presented thus far provide evidence that a superpixel approaches offers better explainability when other mitigating factors are controlled for compared to Grad-CAM. Therefore, this investigation will attempt at applying this approach to resolve current trust-related setbacks that AI in autonomous vehicles (AVs) continue to face.
This aim of this investigating is to understand the extent at which eXplainable Artificial Intelligence (XAI) contributes to autonomous vehicle intelligence accountability, stakeholder trust and regulatory compliance setbacks with connectivity being at the center of research. Exploring an umbrella of XAI vulnerabilities as detailed by (Mohseni et al., 2022) such as interpretability, verifiability and performance limitations that are inherit with XAI in AVs will form the mantle of this research. Understanding the ‘explainability’ dimension is essential to rationalize the inner workings of black-box algorithms, it also adds accountability and transparency dimensions that are of great value to regulators, consumers and service providers (Hussain et al., 2021). This ambiguity in AI decision-making funnels down to the problem of trust with intelligent transportation. Fatal accidents involving autonomous vehicle and automated guided vehicles including Tesla’s autopilot prove that self-driving cars need better models such as true 3D model of their spatial dimensions (Jenssen et al., 2019). This study further concludes that AVs need a better sense of self like humans do and acknowledges the risk in software and sensors leading to this challenge.
Being a critical-mission application, AI-powered vehicles face trust, ethics and transparency issues which hinder their deployment and regulatory acceptance (Kwan et al., 2021). The overarching problem stems from understanding the artificial intelligence responsible for decision-making is not understood by humans due to the lack of expandability of machine learning algorithms (Atakishiyev et al., 2021). Furthermore, the application of black box deep learning models in AVs has divided opinions; mainly between higher accuracy versus interpretability of explainable models. However, for autonomous vehicles to optimally react to near-accident or unavoidable accidents, they need large amounts rare-event data to learn from. Collecting crash data to help with explainable artificial intelligence continues to be a challenge, thus research in XAI based on rare-events has not been fully explored.
References
Atakishiyev, S., Salameh, M., Yao, H., & Goebel, R. (2021). Explainable Artificial Intelligence for Autonomous Driving: A Comprehensive Overview and Field Guide for Future Research Directions. 1–18. http://arxiv.org/abs/2112.11561
Cai, L., Xu, X., Liew, J. H., & Foo, C. S. (2021). Revisiting Superpixels for Active Learning in Semantic Segmentation with Realistic Annotation Costs. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 10983–10992. https://doi.org/10.1109/CVPR46437.2021.01084
Hartley, T., Sidorov, K., Willis, C., & Marshall, D. (2021). SWAG: Superpixels Weighted by average gradients for explanations of CNNs. Proceedings — 2021 IEEE Winter Conference on Applications of Computer Vision, WACV 2021, 423–432. https://doi.org/10.1109/WACV48630.2021.00047
Hussain, F., Hussain, R., & Hossain, E. (2021). Explainable Artificial Intelligence ( XAI ): An Engineering Perspective. 1–11.
Jenssen, G. D., Moen, T., & Johnsen, S. O. (2019). Accidents with Automated Vehicles -Do self-driving cars need a better sense of self ? Accidents with Automated Vehicles — Do self-driving cars need a better sense of self ? October.
Kwan, D., Cysneiros, L. M., & Do Prado Leite, J. C. S. (2021). Towards Achieving Trust Through Transparency and Ethics. Proceedings of the IEEE International Conference on Requirements Engineering, March, 82–93. https://doi.org/10.1109/RE51729.2021.00015
Liu, F., Zhang, X., Wang, H., & Feng, J. (2020). Context-aware superpixel and bilateral entropy-image coherence induces less entropy. Entropy, 22(1), 20. https://doi.org/10.3390/e22010020
Mohseni, S., Yu, Z., Xiao, C., & Yadawa, J. A. Y. (2022). Taxonomy of Machine Learning Safety : A Survey and Primer. 1(1), 1–35.
Suzuki, T. (2020). Superpixel Segmentation Via Convolutional Neural Networks with Regularized Information Maximization. ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing — Proceedings, 2020-May, 2573–2577. https://doi.org/10.1109/ICASSP40776.2020.9054140
Wang, K., Li, L., & Zhang, J. (2020). End-to-end trainable network for superpixel and image segmentation. Pattern Recognition Letters, 140, 135–142. https://doi.org/https://doi.org/10.1016/j.patrec.2020.09.016
Yu, L., & Fan, G. (2021). DrsNet: Dual-resolution semantic segmentation with rare class-oriented superpixel prior. Multimedia Tools and Applications, 80(2), 1687–1706. https://doi.org/10.1007/s11042-020-09691-y
