Image dehazing aims to recover the uncorrupted content from a hazy image. Instead of leveraging traditional low-level or handcrafted image priors as restoration constraints, e.g., dark channels and increased contrast, we propose an end-to-end gated context aggregation network directly restore final haze-free In this network, adopt latest smoothed dilation technique help remove gridding artifacts caused by widely-used dilated convolution with negligible extra parameters, leverage sub-network...

This paper proposes a deep neural network structure that exploits edge information in addressing representative low-level vision tasks such as layer separation and image filtering. Unlike most other learning strategies applied this context, our approach tackles these challenging problems by estimating edges reconstructing images using only cascaded convolutional layers arranged no handcrafted or application-specific image-processing components are required. We apply the resulting...

The emergence of low-cost 3D printers steers the investigation new geometric problems that control quality fabricated object. In this paper, we present a method to reduce material cost and weight given object while providing durable printed model is resistant impact external forces. We introduce hollowing optimization algorithm based on concept honeycomb-cells structure. Honeycombs structures are known be minimal strength in tension. utilize Voronoi diagram compute irregular honeycomb-like...

We introduce JumpCut, a new mask transfer and interpolation method for interactive video cutout. Given source frame which foreground is already available, we compute an estimate of the at another, typically non-successive, target frame. Observing that background regions exhibit different motions, leverage these differences by computing two separate nearest-neighbor fields (split-NNF) from to These NNFs are then used jointly predict coherent labeling pixels in The same split-NNF also aid...

Image smoothing represents a fundamental component of many disparate computer vision and graphics applications. In this paper, we present unified unsupervised (label-free) learning framework that facilitates generating flexible high-quality effects by directly from data using deep convolutional neural networks (CNNs). The heart the design is training signal as novel energy function includes an edge-preserving regularizer which helps maintain important yet potentially vulnerable image...

While invaluable for many computer vision applications, decomposing a natural image into intrinsic reflectance and shading layers represents challenging, underdetermined inverse problem. As opposed to strict reliance on conventional optimization or filtering solutions with strong prior assumptions, deep learning based approaches have also been proposed compute decompositions when granted access sufficient labeled training data. The downside is that current data sources are quite limited,...

In this work, we tackle the problem of category-level online pose tracking objects from point cloud sequences. For first time, propose a unified framework that can handle 9DoF for novel rigid object instances as well per-part articulated known categories. Here pose, comprising 6D and 3D size, is equivalent to amodal bounding box representation with free pose. Given depth at current frame estimated last frame, our end-to-end pipeline learns accurately update Our composed three modules: 1)...

Many different deep networks have been used to approximate, accelerate or improve traditional image operators. Among these operators, many contain parameters which need be tweaked obtain the satisfactory results, we refer as "parameterized operators". However, most existing trained for operators are only designed one specific parameter configuration, does not meet needs of real scenarios that usually require flexible settings. To overcome this limitation, propose a new decoupled learning...

This paper proposes a method for representation learning of multimodal data using contrastive losses. A traditional approach is to contrast different modalities learn the information shared among them. However, that could fail complementary synergies between might be useful downstream tasks. Another concatenate all into tuple and then positive negative correspondences. consider only stronger while ignoring weaker ones. To address these issues, we propose novel objective, TupleInfoNCE. It...

We present C · ASE, an efficient and effective framework that learns Conditional Adversarial Skill Embeddings for physics-based characters. ASE enables the physically simulated character to learn a diverse repertoire of skills while providing controllability in form direct manipulation be performed. This is achieved by dividing heterogeneous skill motions into distinct subsets containing homogeneous samples training low-level conditional model behavior distribution. The skill-conditioned...

Object goal navigation (ObjectNav) in unseen environments is a fundamental task for Embodied AI. Agents existing works learn ObjectNav policies based on 2D maps, scene graphs, or image sequences. Considering this happens 3D space, 3D-aware agent can advance its capability via learning from fine-grained spatial information. However, leveraging representation be prohibitively unpractical policy floor-level task, due to low sample efficiency and expensive computational cost. In work, we propose...

Visual prompt, a pair of before-and-after edited images, can convey indescribable imagery transformations and prosper in image editing. However, current visual prompt methods rely on pretrained text-guided image-to-image generative model that requires triplet text, before, after images for retraining over text-to-image model. Such crafting triplets processes limit the scalability generalization In this paper, we present framework based any single without reliance explicit thus enhancing...

Self-supervised representation learning is a critical problem in computer vision, as it provides way to pretrain feature extractors on large unlabeled datasets that can be used an initialization for more efficient and effective training downstream tasks. A promising approach use contrastive learn latent space where features are close similar data samples far apart dissimilar ones. This has demonstrated tremendous success pretraining both image point cloud extractors, but been barely...

Optical flow in heavy rainy scenes is challenging due to the presence of both rain steaks and veiling effect, which break existing optical constraints. Concerning this, we propose a deep-learning based method designed handle rain. We introduce feature multiplier our network that transforms features an image affected by effect into are less it, call veiling-invariant features. establish new mapping operation space produce streak-invariant The on pyramid structure input images, basic idea...

Autonomous part assembly is a challenging yet crucial task in 3D computer vision and robotics. Analogous to buying an IKEA furniture, given set of parts that can assemble single shape, intelligent agent needs perceive the geometry, reason propose pose estimations for input parts, finally call robotic planning control routines actuation. In this paper, we focus on estimation subproblem from side involving geometric relational reasoning over geometry. Essentially, generative predict 6-DoF...

Point cloud is an important 3D data representation widely used in many essential applications. Leveraging deep neural networks, recent works have shown great success processing point clouds. However, those networks are vulnerable to various adversarial attacks, which can be summarized as two primary types: perturbation that affects local distribution, and surface distortion causes dramatic changes geometry. In this paper, we simultaneously address both the aforementioned attacks by learning...

Localizing the camera in a known indoor environment is key building block for scene mapping, robot navigation, AR, etc. Recent advances estimate pose via optimization over 2D/3D-3D correspondences established between coordinates 2D/3D space and 3D world space. Such mapping estimated with either convolution neural network or decision tree using only static input image sequence, which makes these approaches vulnerable to dynamic environments that are quite common yet challenging real world. To...

We study the problem of multi-robot active mapping, which aims for complete scene map construction in minimum time steps. The key to this lies goal position estimation enable more efficient robot movements. Previous approaches either choose frontier as via a myopic solution that hinders efficiency, or maximize long-term value reinforcement learning directly regress position, but does not guarantee construction. In paper, we propose novel algorithm, namely NeuralCoMapping, takes advantage...

Perceiving and manipulating 3D articulated objects (e.g., cabinets, doors) in human environments is an important yet challenging task for future home-assistant robots. The space of exceptionally rich their myriad semantic categories, diverse shape geometry, complicated part functionality. Previous works mostly abstract kinematic structure with estimated joint parameters poses as the visual representations objects. In this paper, we propose object-centric actionable priors a novel...

Traditional image processing operators often provide some control parameters to tweak the final results. Recently, different convolutional neural networks have been used approximate or improve these operators. However, in those methods, one single model can only handle operator of a specific parameter value and does not support tuning. In this paper, we propose new plugin module, "Adaptive Filterbank Pyramid", which be inserted into backbone network multiple continuous Our module explicitly...