Abstract:We introduce Generalizable 3D-Language Feature Fields (g3D-LF), a 3D representation model pre-trained on large-scale 3D-language dataset for embodied tasks. Our g3D-LF processes posed RGB-D images from agents to encode feature fields for: 1) Novel view representation predictions from any position in the 3D scene; 2) Generations of BEV maps centered on the agent; 3) Querying targets using multi-granularity language within the above-mentioned representations. Our representation can be generalized to unseen environments, enabling real-time construction and dynamic updates. By volume rendering latent features along sampled rays and integrating semantic and spatial relationships through multiscale encoders, our g3D-LF produces representations at different scales and perspectives, aligned with multi-granularity language, via multi-level contrastive learning. Furthermore, we prepare a large-scale 3D-language dataset to align the representations of the feature fields with language. Extensive experiments on Vision-and-Language Navigation under both Panorama and Monocular settings, Zero-shot Object Navigation, and Situated Question Answering tasks highlight the significant advantages and effectiveness of our g3D-LF for embodied tasks.
Abstract:As video generation models advance rapidly, assessing the quality of generated videos has become increasingly critical. Existing metrics, such as Fr\'echet Video Distance (FVD), Inception Score (IS), and ClipSim, measure quality primarily in latent space rather than from a human visual perspective, often overlooking key aspects like appearance and motion consistency to physical laws. In this paper, we propose a novel metric, VAMP (Visual Appearance and Motion Plausibility), that evaluates both the visual appearance and physical plausibility of generated videos. VAMP is composed of two main components: an appearance score, which assesses color, shape, and texture consistency across frames, and a motion score, which evaluates the realism of object movements. We validate VAMP through two experiments: corrupted video evaluation and generated video evaluation. In the corrupted video evaluation, we introduce various types of corruptions into real videos and measure the correlation between corruption severity and VAMP scores. In the generated video evaluation, we use state-of-the-art models to generate videos from carefully designed prompts and compare VAMP's performance to human evaluators' rankings. Our results demonstrate that VAMP effectively captures both visual fidelity and temporal consistency, offering a more comprehensive evaluation of video quality than traditional methods.
Abstract:Understanding robot behaviors and experiences through natural language is crucial for developing intelligent and transparent robotic systems. Recent advancement in large language models (LLMs) makes it possible to translate complex, multi-modal robotic experiences into coherent, human-readable narratives. However, grounding real-world robot experiences into natural language is challenging due to many reasons, such as multi-modal nature of data, differing sample rates, and data volume. We introduce RONAR, an LLM-based system that generates natural language narrations from robot experiences, aiding in behavior announcement, failure analysis, and human interaction to recover failure. Evaluated across various scenarios, RONAR outperforms state-of-the-art methods and improves failure recovery efficiency. Our contributions include a multi-modal framework for robot experience narration, a comprehensive real-robot dataset, and empirical evidence of RONAR's effectiveness in enhancing user experience in system transparency and failure analysis.
Abstract:Retrieval-augmented generation (RAG) has gained wide attention as the key component to improve generative models with external knowledge augmentation from information retrieval. It has shown great prominence in enhancing the functionality and performance of large language model (LLM)-based applications. However, with the comprehensive application of RAG, more and more problems and limitations have been identified, thus urgently requiring further fundamental exploration to improve current RAG frameworks. This workshop aims to explore in depth how to conduct refined and reliable RAG for downstream AI tasks. To this end, we propose to organize the first R3AG workshop at SIGIR-AP 2024 to call for participants to re-examine and formulate the basic principles and practical implementation of refined and reliable RAG. The workshop serves as a platform for both academia and industry researchers to conduct discussions, share insights, and foster research to build the next generation of RAG systems. Participants will engage in discussions and presentations focusing on fundamental challenges, cutting-edge research, and potential pathways to improve RAG. At the end of the workshop, we aim to have a clearer understanding of how to improve the reliability and applicability of RAG with more robust information retrieval and language generation.
Abstract:Road ponding, a prevalent traffic hazard, poses a serious threat to road safety by causing vehicles to lose control and leading to accidents ranging from minor fender benders to severe collisions. Existing technologies struggle to accurately identify road ponding due to complex road textures and variable ponding coloration influenced by reflection characteristics. To address this challenge, we propose a novel approach called Self-Attention-based Global Saliency-Enhanced Network (AGSENet) for proactive road ponding detection and traffic safety improvement. AGSENet incorporates saliency detection techniques through the Channel Saliency Information Focus (CSIF) and Spatial Saliency Information Enhancement (SSIE) modules. The CSIF module, integrated into the encoder, employs self-attention to highlight similar features by fusing spatial and channel information. The SSIE module, embedded in the decoder, refines edge features and reduces noise by leveraging correlations across different feature levels. To ensure accurate and reliable evaluation, we corrected significant mislabeling and missing annotations in the Puddle-1000 dataset. Additionally, we constructed the Foggy-Puddle and Night-Puddle datasets for road ponding detection in low-light and foggy conditions, respectively. Experimental results demonstrate that AGSENet outperforms existing methods, achieving IoU improvements of 2.03\%, 0.62\%, and 1.06\% on the Puddle-1000, Foggy-Puddle, and Night-Puddle datasets, respectively, setting a new state-of-the-art in this field. Finally, we verified the algorithm's reliability on edge computing devices. This work provides a valuable reference for proactive warning research in road traffic safety.
Abstract:This study presents the first implementation of multilayer neural networks on a memristor/CMOS integrated system on chip (SoC) to simultaneously detect multiple diseases. To overcome limitations in medical data, generative AI techniques are used to enhance the dataset, improving the classifier's robustness and diversity. The system achieves notable performance with low latency, high accuracy (91.82%), and energy efficiency, facilitated by end-to-end execution on a memristor-based SoC with ten 256x256 crossbar arrays and an integrated on-chip processor. This research showcases the transformative potential of memristive in-memory computing hardware in accelerating machine learning applications for medical diagnostics.
Abstract:As large language models (LLMs) are rapidly advancing and achieving near-human capabilities, aligning them with human values is becoming more urgent. In scenarios where LLMs outperform humans, we face a weak-to-strong alignment problem where we need to effectively align strong student LLMs through weak supervision generated by weak teachers. Existing alignment methods mainly focus on strong-to-weak alignment and self-alignment settings, and it is impractical to adapt them to the much harder weak-to-strong alignment setting. To fill this gap, we propose a multi-agent contrastive preference optimization (MACPO) framework. MACPO facilitates weak teachers and strong students to learn from each other by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. To get this, we devise a mutual positive behavior augmentation strategy to encourage weak teachers and strong students to learn from each other's positive behavior and further provide higher quality positive behavior for the next iteration. Additionally, we propose a hard negative behavior construction strategy to induce weak teachers and strong students to generate familiar negative behavior by fine-tuning on negative behavioral data. Experimental results on the HH-RLHF and PKU-SafeRLHF datasets, evaluated using both automatic metrics and human judgments, demonstrate that MACPO simultaneously improves the alignment performance of strong students and weak teachers. Moreover, as the number of weak teachers increases, MACPO achieves better weak-to-strong alignment performance through more iteration optimization rounds.
Abstract:Deep neural networks (DNNs) at convergence consistently represent the training data in the last layer via a highly symmetric geometric structure referred to as neural collapse. This empirical evidence has spurred a line of theoretical research aimed at proving the emergence of neural collapse, mostly focusing on the unconstrained features model. Here, the features of the penultimate layer are free variables, which makes the model data-agnostic and, hence, puts into question its ability to capture DNN training. Our work addresses the issue, moving away from unconstrained features and studying DNNs that end with at least two linear layers. We first prove generic guarantees on neural collapse that assume (i) low training error and balancedness of the linear layers (for within-class variability collapse), and (ii) bounded conditioning of the features before the linear part (for orthogonality of class-means, as well as their alignment with weight matrices). We then show that such assumptions hold for gradient descent training with weight decay: (i) for networks with a wide first layer, we prove low training error and balancedness, and (ii) for solutions that are either nearly optimal or stable under large learning rates, we additionally prove the bounded conditioning. Taken together, our results are the first to show neural collapse in the end-to-end training of DNNs.
Abstract:Recent advancements in 3D Gaussian Splatting (3D-GS) have revolutionized novel view synthesis, facilitating real-time, high-quality image rendering. However, in scenarios involving reflective surfaces, particularly mirrors, 3D-GS often misinterprets reflections as virtual spaces, resulting in blurred and inconsistent multi-view rendering within mirrors. Our paper presents a novel method aimed at obtaining high-quality multi-view consistent reflection rendering by modelling reflections as physically-based virtual cameras. We estimate mirror planes with depth and normal estimates from 3D-GS and define virtual cameras that are placed symmetrically about the mirror plane. These virtual cameras are then used to explain mirror reflections in the scene. To address imperfections in mirror plane estimates, we propose a straightforward yet effective virtual camera optimization method to enhance reflection quality. We collect a new mirror dataset including three real-world scenarios for more diverse evaluation. Experimental validation on both Mirror-Nerf and our real-world dataset demonstrate the efficacy of our approach. We achieve comparable or superior results while significantly reducing training time compared to previous state-of-the-art.
Abstract:3D Gaussian Splatting (3DGS) integrates the strengths of primitive-based representations and volumetric rendering techniques, enabling real-time, high-quality rendering. However, 3DGS models typically overfit to single-scene training and are highly sensitive to the initialization of Gaussian ellipsoids, heuristically derived from Structure from Motion (SfM) point clouds, which limits both generalization and practicality. To address these limitations, we propose GS-Net, a generalizable, plug-and-play 3DGS module that densifies Gaussian ellipsoids from sparse SfM point clouds, enhancing geometric structure representation. To the best of our knowledge, GS-Net is the first plug-and-play 3DGS module with cross-scene generalization capabilities. Additionally, we introduce the CARLA-NVS dataset, which incorporates additional camera viewpoints to thoroughly evaluate reconstruction and rendering quality. Extensive experiments demonstrate that applying GS-Net to 3DGS yields a PSNR improvement of 2.08 dB for conventional viewpoints and 1.86 dB for novel viewpoints, confirming the method's effectiveness and robustness.