CleverDistiller: Simple and Spatially Consistent Cross-modal Distillation

The study introduces the Large Shape and Textures dataset (LAS&T), a comprehensive collection of diverse shapes and textures extracted from natural images. This dataset is utilized to evaluate the performance of leading Large Vision-Language Models (VLMs) in recognizing and representing shapes and textures in various contexts. Results indicate that VLMs still lag behind human capabilities in shape recognition, particularly when variations in orientation, texture, and color are present.

LiSTAR is a novel generative world model designed to synthesize high-fidelity 4D LiDAR data for autonomous driving applications. It addresses challenges such as the sensor's spherical geometry and temporal sparsity of point clouds by introducing a Hybrid-Cylindrical-Spherical representation and a Spatio-Temporal Attention mechanism. This approach enhances data fidelity and temporal coherence, making it a significant advancement in creating scalable simulation environments for autonomous vehicles.

The paper presents Mem-MLP, a novel method for real-time 3D human motion generation from sparse inputs using a multi-layer perceptron (MLP) model. This approach addresses the limitations of existing AR/VR systems that primarily track head and hands, resulting in incomplete full-body reconstructions. Mem-MLP incorporates a Memory-Block to represent missing sensor data and improves temporal consistency through multi-task learning, enhancing accuracy in motion tracking.

The paper discusses dataset distillation, aiming to create a small set of synthetic images that can train a model to match the performance of one trained on a larger dataset. Unlike previous methods that focus on randomly initialized models, this research targets pre-trained self-supervised vision models. The proposed Linear Gradient Matching method optimizes synthetic images to produce similar gradients in a linear classifier as real data, enhancing the training process.

Self-supervised depth estimation has become crucial in fields like autonomous driving and robotics. However, existing methods struggle in adverse weather conditions, leading to performance degradation. To tackle this, a new framework called SEC-Depth is proposed, which utilizes intermediate training parameters to create evolving latency models. This approach aims to enhance depth estimation robustness under challenging conditions through a self-evolution contrastive learning scheme.

Split DNNs facilitate edge devices by transferring heavy computations to cloud servers, but this approach raises privacy concerns as intermediate features can be exploited to reconstruct private inputs through Feature Inversion Attacks (FIA). Current FIA methods yield limited reconstruction quality, complicating the assessment of privacy risks. The introduction of FIA-Flow, a black-box FIA framework, enhances image reconstruction fidelity using a Latent Feature Space Alignment Module and Deterministic Inversion Flow Matching.

The paper titled 'Scriboora: Rethinking Human Pose Forecasting' evaluates various algorithms for predicting human poses based on past observations. It highlights reproducibility issues and introduces a unified training and evaluation pipeline. The study demonstrates that recent speech models can be adapted to enhance pose forecasting performance, and assesses model robustness using noisy joint coordinates to better reflect real-world applications.

The paper presents a novel physical adversarial attack targeting stereo matching models used in autonomous driving. Unlike traditional attacks that utilize 2D patches, this approach employs a 3D physical adversarial example (PAE) with global camouflage texture, enhancing visual consistency across various viewpoints. Additionally, a new 3D stereo matching rendering module is introduced to align the PAE with real-world positions in binocular vision, addressing the disparity effects of stereo cameras.