Text representation plays a critical role in tasks like clustering, retrieval, and other downstream applications. With the emergence of large language models (LLMs), there is increasing interest in harnessing their capabilities for this purpose. However, most of the LLMs are inherently causal and optimized for next-token prediction, making them suboptimal for producing holistic representations. To address this, recent studies introduced pretext tasks to adapt LLMs for text representation. Most of these tasks, however, rely on token-level prediction objectives, such as the masked next-token prediction (MNTP) used in LLM2Vec. In this work, we explore the untapped potential of context compression as a pretext task for unsupervised adaptation of LLMs. During compression pre-training, the model learns to generate compact memory tokens, which substitute the whole context for downstream sequence prediction. Experiments demonstrate that a well-designed compression objective can significantly enhance LLM-based text representations, outperforming models trained with token-level pretext tasks. Further improvements through contrastive learning produce a strong representation model (LLM2Comp) that outperforms contemporary LLM-based text encoders on a wide range of tasks while being more sample-efficient, requiring significantly less training data.

Semantic mapping based on information of objects has become a crucial component for the surrounding comprehension and the more robust navigation. In this paper, we propose a system for simultaneous localization and mapping (SLAM) that combines multiple objects tracking and factor graph optimization with semantically meaningful landmarks to achieve accurate monocular semantic mapping. Firstly, the process of object recognition uses a vanishing point sampling-based approach to efficiently infer the class and position of object landmarks from 2D bounding box object detection. Secondly, The semantic frontend utilizes local matching-based data association to track raw cuboid proposals. It can provide semantic constraints to reduce cuboid scale drift and improve its position estimation. Finally, we present a multi-view factor graph optimization which can use motion modal of the camera to optimize stable cuboids. The semantic mapping experiments on our own built virtual scene show better accuracy and robustness over existing approaches. We evaluate the effectiveness of our approach on public KITTI datasets and a real scene.

This paper presents a direct visual odometry method with adaptive stereo coupling factor and relocalization. The adaptive stereo coupling factor is calculated by considering the number of co-visibility frames, which makes stereo direct odometry more robust. The map is built up with keyframes, connecting relationship of keyframes, poses of keyframes and descriptors. The map will be saved automatically when the proposed system shuts down, and will be loaded automatically when the proposed system starts up. Our proposed system always builds a new map until a loop closure is found between new map and loaded map. When loop closure in a loaded map is found, the proposed system will locate camera in the loaded map. Moreover, our proposed system detects loop closure with feature-based bag-of-words (BoW) method on a low resolution level, which makes the loop closure detection achieve realtime. Evaluation results on the KITTI dataset show that, with the aid of the adaptive stereo coupling factor, the stereo visual odometry becomes more accurate and robust. And evaluation results in real world show that the proposed system can successfully detect loop closure and relocate in realtime.

This article proposes a hardware accelerator for image feature extraction based on the oriented features from accelerated segment test (FAST) and rotated binary robust independent elementary features (ORB) algorithm. The architecture adopts a hybrid workflow to process three scales in parallel as an octave, and multiple octaves are processed in serial by time-sharing the DRAM. In order to support any image resolution, a block-wise dataflow is used on the serial-parallel architecture. The overlapped data between two blocks is reused. As a result, the on-chip memory is limited to 1.47 Mb, and the DRAM bandwidth is compressed by 33%. Besides, 3×3 nonmaximum suppression with heap sorting is applied to balance keypoint distribution in 2-D. It improves the valid keypoint match ratio by 7.09%. In addition, parallel processing in keypoint detection increases 2× throughout. Moreover, approximate computing and superscalar processing reduce the timing cost in orientation estimation and descriptor generation by 92.78% and 67% respectively. Compared to the non-optimized baseline architecture, the proposed architecture saves the total timing cost by 82.4%. The accelerator is implemented in a Xilinx MPSoC, and it achieves 108 fps on full-HD images at 200 MHz under 873 mW.

In the field of autonomous vehicles and robotics research, a well-designed testbed can provide a convenient and safe development environment. In this paper, we propose and implement a reconfigurable testbed equipped with multiple heterogeneous sensors and a compact but powerful local computation unit. The proposed testbed benefits from a self-sustaining design, and can be smoothly reconfigured to different varieties of vehicles. These features make the testbed a novel and ideal platform for testing and verifying navigation, recognition and control algorithms under diverse scenarios.