The multimodal deep neural networks, represented by CLIP, have generated rich downstream applications owing to their excellent performance, thus making understanding the decision-making process of CLIP an essential research topic. Due to the complex structure and the massive pre-training data, it is often regarded as a black-box model that is too difficult to understand and interpret. Concept-based models map the black-box visual representations extracted by deep neural networks onto a set of human-understandable concepts and use the concepts to make predictions, enhancing the transparency of the decision-making process. However, these methods involve the datasets labeled with fine-grained attributes by expert knowledge, which incur high costs and introduce excessive human prior knowledge and bias. In this paper, we observe the long-tail distribution of concepts, based on which we propose a two-stage Concept Selection Model (CSM) to mine core concepts without introducing any human priors. The concept greedy rough selection algorithm is applied to extract head concepts, and then the concept mask fine selection method performs the extraction of core concepts. Experiments show that our approach achieves comparable performance to end-to-end black-box models, and human evaluation demonstrates that the concepts discovered by our method are interpretable and comprehensible for humans.

Statistical learning techniques and increased computational power have facilitated the development of self-driving car technology. However, a limiting factor has been the high expense of scaling and maintaining high-definition (HD) maps. These maps are a crucial backbone for many approaches to self-driving technology. In response to this challenge, we present an approach that fuses pre-built point cloud map data with images to automatically and accurately identify static landmarks such as roads, sidewalks, and crosswalks. Our pipeline utilizes semantic segmentation of 2D images, associates semantic labels with points in point cloud maps to pinpoint locations in the physical world, and employs a confusion matrix formulation to generate a probabilistic bird's-eye view semantic map from semantic point clouds. The approach has been tested in an urban area with different segmentation networks to generate a semantic map with road features. The resulting map provides a rich context of the environment that is valuable for downstream tasks such as trajectory generation and intent prediction. Moreover, it has the potential to be extended to the automatic generation of HD maps for semantic features. The entire software pipeline is implemented in the robot operating system (ROS), a widely used robotics framework, and made available.

This work explores methodologies for dynamic trajectory generation for urban driving environments by utilizing coarse global plan representations. In contrast to state-of-the-art architectures for autonomous driving that often leverage lane-level high-definition (HD) maps, we focus on minimizing required map priors that are needed to navigate in dynamic environments that may change over time. To incorporate high-level instructions (i.e., turn right vs. turn left at intersections), we compare various representations provided by lightweight and open-source OpenStreetMaps (OSM) and formulate a conditional generative model strategy to explicitly capture the multimodal characteristics of urban driving. To evaluate the performance of the models introduced, a data collection phase is performed using multiple full-scale vehicles with ground truth labels. Our results show potential use cases in dynamic urban driving scenarios with real-time constraints. The dataset is released publicly as part of this work in combination with code and benchmarks.