Image dehazing is a representative low-level vision task that estimates latent haze-free images from hazy images. In recent years, convolutional neural network-based methods have dominated image dehazing. However, Transformers, which has recently made breakthrough in high-level tasks, not brought new dimensions to We start with the popular Swin Transformer and find several of its key designs are unsuitable for To this end, we propose DehazeFormer, consists various improvements, such as...

In autonomous driving, deep models have shown remarkable performance across various visual perception tasks with the demand of high-quality and huge-diversity training datasets. Such datasets are expected to cover driving scenarios adverse weather, lighting conditions diverse moving objects. However, manually collecting these data presents huge challenges expensive cost. With rapid development large generative models, we propose DriveDiTFit, a novel method for efficiently generating Driv ing...

Recently, the decentralized optimization problem is attracting growing attention. Most existing methods are deterministic with high per-iteration cost and have a convergence rate quadratically depending on condition number. Besides, dense communication necessary to ensure even if dataset sparse. In this paper, we generalize monotone operator root finding problem, propose stochastic algorithm named DSBA that (i) converges geometrically linearly number, (ii) can be implemented using sparse...

Image enhancement is a subjective process whose targets vary with user preferences. In this paper, we propose deep learning-based image method covering multiple tonal styles using only single model dubbed StarEnhancer. It can transform an from one style to another, even if that unseen. With simple one-time setting, users customize the make enhanced images more in line their aesthetics. To practical, well-designed enhancer 4K-resolution over 200 FPS but surpasses contemporaneous methods terms...

Tactile sensation plays a crucial role in the development of multi-modal large models and embodied intelligence. To collect tactile data with minimal cost as possible, series studies have attempted to generate images by vision-to-touch image translation. However, compared text modality, visual modality-driven generation cannot accurately depict human sensation. In this work, we analyze characteristics detail from two granularities: object-level (tactile texture, shape), sensor-level (gel...

This paper proposes a new approach to upsample depth maps when aligned high-resolution color images are given. Such task is referred as guided upsampling in our work. We formulate this problem based on the recently developed sparse representation analysis models. More specifically, we exploit cosparsity of analytic operators performed map, together with data fidelity and smoothness constraints for upsampling. The formulated solved by greedy pursuit algorithm. Since relies such Wavelet...

Wearable cameras, such as Google Glass and Go Pro, enable video data collection over larger areas from different views. In this paper, we tackle a new problem of locating the co-interest person (CIP), i.e., one who draws attention most camera wearers, temporally synchronized videos taken by multiple wearable cameras. Our basic idea is to exploit motion patterns people use them correlate persons across videos, instead performing appearance-based matching in traditional...

Motion detection surveillance technology came about as a relief for the generally time-consuming reviewing process that normal video system offers. It has gained lot of interests over past few years. In this paper, we propose motion system, consisting its graphic user interface (GUI) and method detection, through study evaluation currently available products methods. The proposed is efficient convenient both office home uses.

We improve inverse reinforcement learning (IRL) by applying dimension reduction methods to automatically extract abstract features from human-demonstrated policies, deal with the cases where are either unknown or numerous. The importance rating of each feature is incorporated into reward function. Simulation performed on a task driving in five-lane highway, controlled car has largest fixed speed among all cars. Performance almost 10.6% better average than without ratings.

In this paper, we propose the first fully push-forward-based Distributional Reinforcement Learning algorithm, called Push-forward-based Actor-Critic EncourageR (PACER). Specifically, PACER establishes a stochastic utility value policy gradient theorem and simultaneously leverages push-forward operator in construction of both actor critic. Moreover, based on maximum mean discrepancies (MMD), novel sample-based encourager is designed to incentivize exploration. Experimental evaluations various...

Particle-based Variational Inference (ParVI) methods approximate the target distribution by iteratively evolving finite weighted particle systems. Recent advances of ParVI reveal benefits accelerated position update strategies and dynamic weight adjustment approaches. In this paper, we propose first framework that possesses both dynamical simultaneously, named General Accelerated Dynamic-Weight (GAD-PVI) framework. Generally, GAD-PVI simulates semi-Hamiltonian gradient flow on a novel...

Continual Learning (CL) aims to learn in non-stationary scenarios, progressively acquiring and maintaining knowledge from sequential tasks. Recent Prompt-based (PCL) has achieved remarkable performance with Pre-Trained Models (PTMs). These approaches grow a prompt sets pool by adding new set of prompts when learning each task (\emph{prompt learning}) adopt matching mechanism select the correct for testing sample retrieval}). Previous studies focus on latter stage improving enhance Prompt...

In order to solve the problem of environment generalization in continuous state space, an obstacle avoidance method based on region location is proposed. The divided into three steps: (1) Using Region Proposal Network (RPN) localize area; (2). map established by regional position mapping relation; and (3). Deep Q-Learning (DQN) used realize collision detection robot, then pixel module introduced finally simulation distance sensor combined obtain between robot whether or not. this paper,...

We present a novel system for visual based contour tracking and apply it to track the movement of human's head. Currently available technology is used in implementation take care low level video processing acquisition. The consists two main parts. First, there an error sensitive Gaussian particle filter. It estimates posterior density shape being tracked by modeling distribution as collection weighted samples. second part observer which takes estimate current position, uses image features...

Because of the complexity outdoor terrain and variety robot missions, autonomous mobile robots (AMR) have to carry out missions in unstructured impossibly predicted environment. The acquisition accurate environment parameters is crucial for path planning. In this article, idea mission-plan introduced on basis plan. whole mission decomposed into several sub-missions during modeling process. And one case-based mission-planning method (CBMP) applied filter optimize robot's geometric kinetic...

Wearable cameras, such as Google Glass and Go Pro, enable video data collection over larger areas from different views. In this paper, we tackle a new problem of locating the co-interest person (CIP), i.e., one who draws attention most camera wearers, temporally synchronized videos taken by multiple wearable cameras. Our basic idea is to exploit motion patterns people use them correlate persons across videos, instead performing appearance-based matching in traditional...

Both improper initialization and fake Gaussian components are critical problems in GMM-based foreground detection. The former can lead to a poor local maximum, while the latter invokes unhandled disturbance. To eliminate these destructive impacts, two kinds of feedback knowledge introduced: positive negative prior. For appropriate initialization, high level modules provide prior informations by outlining rough objects using optical flow. Moreover, evidences form Dirichlet distribution...