Transfer Learning for Human & AI

DA-VSN is a domain adaptive video segmentation network that addresses domain gaps in videos by temporal consistency regularization (TCR) for consecutive frames of target-domain videos . The network is based on cross-domain TCR that guides the prediction of target frames to have similar temporal consistency as that of source frames (learnt from annotated source data) viaadversarial learning .…

In this paper we present SurfaceNet, an approach for estimatingspatially-varying bidirectional reflectance distribution function (SVBRDF)material properties from a single image . We pose the problem as an imagetranslation task and propose a novel patch-based generative adversarial network that is able to produce high-quality, high-resolution surface reflectancemaps .…

This document describes the generalized moving peaks benchmark (GMPB) and how it can be used to generate problem instances for continuous large-scale dynamic optimization problems . It presents a set of 15 benchmark problems, the relevantsource code, and a performance indicator .…

Many quantum (convolutional) circuit ansaetze are proposed forgrayscale images classification tasks with promising empirical results . But the intra-channel information that is useful for vision tasks is not extracted effectively . This is the first work of a quantum convolutional circuit to deal with RGB images with a higher test accuracy compared to the purely classical CNNs .…

SuperCaustics is a real-time,open-source simulation of transparent objects designed for deep learning applications . It uses hardware ray-tracing to support caustics, dispersion, and refraction . The neural network achieved performance comparable to the state-of-the-art on a real world dataset using only 10% of the training data and in a fraction of training time .…

This paper introduces the sixth Oriental Language Recognition (OLR) 2021Challenge . It intends to improve the performance of language recognitionsystems and speech recognition systems within multilingual scenarios . The dataprofile, four tasks, two baselines, and the evaluation principles are presented .…

Forest carbon offsets are increasingly popular and can play a significant role in financing climate mitigation, forest conservation, and reforestation . Measuring how much carbon is stored in forests is, however, still largely done via expensive, time-consuming, and sometimes unaccountable field measurements .…

Differentially Private Stochastic Convex Optimization(DP-SCO) has been extensively studied in recent years . Most of the previous work can only handle either regular data distribution or irregulardata in the low dimensional space case . To better understand the challengesarising from irregular data distribution, in this paper, we provide the firststudy on the problem with heavy-tailed data in the high dimensionalspace .…

We propose a novel method for solving regression tasks using few-shot or weaksupervision . At the core of our method is the fundamental observation that GANsare incredibly successful at encoding semantic information within their latentspace, even in a completely unsupervised setting .…

Image inpainting aims to complete the missing or corrupted regions of images with realistic contents . WaveFill decomposes images by using discrete wavelettransform (DWT) that preserves spatial information naturally . It applies L1reconstruction loss to the decomposed low-frequency bands and adversarial loss .…

Density estimation plays a crucial role in many data analysis tasks . It is used in tasks as diverse as analyzing population data, spatiallocations in 2D sensor readings, or reconstructing scenes from 3D scans . In this paper, we introduce a learned, data-driven deep density estimation (DDE)to infer PDFs in an accurate and efficient manner, while being independent of domain dimensionality or sample size .…

Recent state-of-the-art active learning methods have mostly leveragedGenerative Adversarial Networks (GAN) for sample acquisition . However, GAN isusually known to suffer from instability and sensitivity to hyper-parameters . In contrast to these methods, we propose in this paper a novel active learningframework that we call Maximum Classifier Discrepancy for Active Learning(MCDAL) which takes the prediction discrepancies between multiple classifiers .…

Future urban transportation concepts include a mixture of ground and air vehicles with varying degrees of autonomy in a congested environment . Safe and efficient transportation requires reasoning about the 3Dflow of traffic and properly modeling uncertainty . This paper explores a Bayesian approach that captures our uncertainty in the map given training data .…

Feature pyramids and iterative refinement have recently led to great progress in optical flow estimation . However, downsampling in feature pyramids can cause blending of foreground objects with background . We propose anovel Residual Feature Pyramid Module (RFPM) which retains important details inthe feature map without changing the overall design of the overall iterative .…

A novel shape-perception network (HSPN) is proposed to reconstruct the 3D point clouds (PCs) of specific brains from one single incomplete image with low latency . 3D shape reconstruction is essential in the navigation of minimally-invasive and auto robot-guided surgeries whose operating environments are indirect and narrow .…

Existing Unsupervised Domain Adaptation (UDA) literature adopts the covariateshift and conditional shift assumptions, which essentially encourage models to learn common features across domains . However, due to the lack of supervision in the target domain, they suffer from the semantic loss .…

Aligned Disentangling Generative AdversarialNetwork (AD-GAN) introduces representationdisentanglement to separate content representation from style representation . With this framework, spatial structure can be preserved explicitly, enabling asignificant reduction of macro-level lossy transformation . AD-GAN leads to significant improvement over the current best unsupervised methods by an average 17.8% relatively (w.r.t.…

Unrealistic feature suppression (UFS) module keeps high-quality features and suppresses unrealistic features . UFS module keeps the training stability of networks and improves the quality ofgenerated images . We demonstrate effectiveness of the module on various models such as WGAN-GP, SNGAN, and BigGAN .…

Recent state-of-the-art active learning methods have mostly leveragedGenerative Adversarial Networks (GAN) for sample acquisition . However, GAN isusually known to suffer from instability and sensitivity to hyper-parameters . In contrast to these methods, we propose in this paper a novel active learningframework that we call Maximum Classifier Discrepancy for Active Learning(MCDAL) which takes the prediction discrepancies between multiple classifiers .…

Identifying unexpected objects on roads in semantic segmentation is crucial in safety-critical applications . Existing approaches use images of unexpected objects from external datasets or require additional training . We propose a simple yet effective approach that standardizes the max logits in order to align the different distributions and reflect the relative meanings of each predicted class .…

A paper develops a new shape model that allows synthesizing noveldistance views by optimizing a continuous signed directional distance function . Unlike an SDF, an SDDF measures distance in a given direction . This allows us to form a shape model without 3D shape supervision, using only distancemeasurements, readily available from depth camera or Lidar sensors .…

Value iteration is a well-known method of solving Markov Decision Processes . However, the computational cost of value iteration quickly becomesfeasible as the size of the state space increases . In this paper, we propose an intuitive algorithm for solving MDPsthat reduces the cost of updates by dynamically grouping together states with similar cost-to-go values .…

Surgical reports aimed at understanding inrobot-assisted surgery can contribute to documenting entry tasks and post-operative analysis . Despite the impressive outcome, the deep learningmodel degrades the performance when applied to different domains encountering domain shifts . In this work, we proposeclass-incremental domain adaptation (CIDA) with a multi-layer transformer-basedmodel to tackle the new classes and domain shift in the target domain to generate surgical reports .…

Explainability of deep neural networks is one of the most challenging and interesting problems in the field . We show that such an approach can decompose the highly convoluted latent latent spaces of registration pipelines in an orthogonal space with several interesting properties .…

Heatmap-based methods dominate in the field of human pose estimation by modelling the output distribution through likelihood heatmaps . Residual Log-likelihood Estimation(RLE) is effective, efficient and flexible . Compared to the conventional regression paradigm, regression with RLE bring 12.4 mAPimprovement on MSCOCO without any test-time overhead .…

A new method that amends the label distribution of each facial image byleveraging correlations among expressions in the semantic space . By utilizing label distribution learning, a probability distribution is assigned for a facial image to express a compound emotion, which effectivelyimproves the problem of label uncertainties and noises occurred in one-hotlabels .…

The AAGC-LSTM combines spatial and temporal dependency in a single network operation . This is made possible by equipping graph convolutions with adjacency adaptivity, which allows for learning unknown dependencies of the human body joints . Tofurther boost accuracy, we propose longitudinal loss weighting to considernatural movement patterns, as well as body-aware contralateral dataaugmentation .…

Deep learning-based 3D object detection has achieved unprecedented success with the advent of large-scale autonomous driving datasets . However, drasticperformance degradation remains a critical challenge for cross-domain deployment . We propose anovel and unified framework, Multi-Level Consistency Network (MLC-Net), whichemploys a teacher-student paradigm to generate adaptive and reliablepseudo-targets .…

Language instruction plays an essential role in the natural language groundednavigation tasks . We propose a DynamicReinforced Instruction Attacker (DR-Attacker) which learns to mislead thenavigator to move to the wrong target by destroying the most instructive information in instructions at different timesteps .…

Current algorithms share one limitation: They rely on directly visible objects . This is a major drawback compared to human behavior, where indirect visual cues caused by the actual object (e.g.,shadows) are already used intuitively to retrieve information . Humans already process light artifacts caused by oncoming vehicles to assume their future appearance, whereas current objectdetection systems rely on the oncoming vehicle’s direct visibility .…

Unsupervised Domain Adaptation for semantic segmentation has gained immensepopularity since it can transfer knowledge from simulation to real (Sim2Real) In this work, we present a novel two-phase adaptation scheme . In the first step, we exhaustively distill source domain knowledge using supervised loss functions .…

We develop a novel data-driven approach to modeling the atmospheric boundarylayer . This approach leads to a nonlocal, anisotropic synthetic turbulencemodel which we refer to as the deep rapid distortion (DRD) model . After calibration, the model can be used to generate syntheticturbulent velocity fields .…

Scene text editing (STE) is a challenging task due to a complex intervention between text and style . We propose a novel representational learning-based STE model that employs textual information as well as visual information . Our experiments demonstrate that RewriteNet achieves betterquantitative and qualitative performance than other comparisons .…

We present four different robust transfer learning strategies for robust mobile scene recognition . Fine-Tuning in combination withextensive data augmentation improves accuracy and robustness in mobile robot place recognition . We achieved state-of-the-art results using variousbaseline convolutional neural networks and showed the robustness againstlighting and viewpoint changes in challenging mobile robot places recognition .…

DistributionAlignment and Random Sampling (DARS) method to produce unbiased pseudo labelsthat match the true class distribution estimated from the labeled data . Code will be available athttps://://github.com/CVMI-Lab/DARS . Experiments on both Cityscapes and PASCAL VOC 2012 datasets demonstrate effectiveness of our approach .…

Single-photon cameras (SPCs) are an emergingsensing modality that are capable of capturing images with high sensitivity . Despite having minimal read-noise, images captured by SPCs in photon-starved conditions still suffer from strong shot noise . We propose photon scale-space a collection of high-SNR imagesspanning a wide range of photons-per-pixel (PPP) levels as guides to train inference model on low photon flux images .…

Semantic segmentation on 3D point clouds is an important task for 3D sceneunderstanding . We train a semantic point cloud segmentation network with only asmall portion of points being labeled . We argue that we can better utilize thelimited supervision information as we densely propagate the supervision signalfrom the labeled points to other points within and across the input samples .…

The method significantly compressesneural network weights to a sparse ternary of [-1,0,1 . It can compress aResNet-18 model from 46 MB to 955KB and a ResNet-50 model from 99 MB to 3.3MB (~30x) The top-1 accuracy on ImageNet drops slightly from 69.7% to65.3% and from 76.15% to 74.47% .…

Deep generative models have made great progress in synthesizing images with human poses and transferring poses of one person to others . Most existing approaches explicitly leverage the pose information extracted from the source images as a conditional input for the generative networks .…

Discontinuous Galerkin (DG)Implicit-Explicit Runge Kutta schemes and LinearMultistep Methods based on Backward-Finite-Differences . The new methods are validated considering two-dimensional benchmark test cases typically used in the fluid dynamics community . A prototype engineering problem consisting of a supersonic flow around a NACA 0012 airfoil with space-time-dependent boundary conditions is presented for which the pressure coefficients are measured .…

A common approach to improve software quality is to use programming guidelines to avoid common kinds of errors . In this paper, we consider theproblem of enforcing guidelines for Featherweight Java (FJ) We formalize guidelines as sets of finite or infinite execution traces and develop aregion-based type and effect system for FJ that can enforce such guidelines .…

The paper proposes a novel approach for gray scale images segmentation . It is based on multiple features extraction from single feature per image pixel, using Echo state network . The newly extractedfeatures — reservoir equilibrium states — reveal hidden image characteristicsthat improve its segmentation via a clustering algorithm .…

Deep learning models have gained great popularity in statistical modeling . Theadvantage of deep learning models is that their solutions are difficult tointerpret and explain . We propose a new network architecture that sharessimilar features as generalized linear models, but provides superior predictivepower benefiting from the art of representation learning .…

This paper explores the convenience of using additional signals apart fromelectroencephalograms . The best overall model, an ensemble of Depth-wiseSeparational Convolutional Neural Networks, has achieved an accuracy of 86.06\% with a Cohen’s Kappa of 0.80 and a $F_{1}$ of $F_1 of $1.77 .…

Skyline plays a pivotal role in mountainous visual geo-localization and localization/navigation of planetary rovers/UAVs and virtual/augmented realityapplications . We present a novel mountainous skyline detection approach wherewe adapt a shallow learning approach to learn a set of filters to discriminate between edges belonging to sky-mountain boundary and others coming from different regions .…

We present a four-pronged approach to build firefighter’s situationalawareness for the first time in the literature . We construct a series of deeplearning frameworks built on top of one another to enhance the safety, efficiency, and successful completion of rescue missions conducted by firefighters in emergency first response settings .…

By lumping all dangling nodes of a web graph into a single node, PageRankcomputation can be proceeded with a much smaller matrix on the case that thenumber of dangling nodes is large . Thus, it saves many computational cost and operations .…

DeepHead is a fully data-driven implementation of the paradigm proposed in the context of brain imaging . It infers the dielectric distribution of the brain at a desired single frequency while making use of an input that spreads over awide band of frequencies .…

A paper uses distanceregularized level set (DRL SE) to explore the segmentation effect of epicardiumand endocardium . Five CT images are used to verify the proposedmethod, and image quality evaluation indexes such as dice score and Hausdorffdistance are used . The results showed that the researchers could separate the inner and outer membrane very well (endocardiumdice = 0.9253, Hausorfff = 7.8740) and epicocardium Hausdice= 0.9687 .…

Avatar based Sign Language Production (SLP) builds up animation from sequences of handmotions, shapes and facial expressions . We propose a novel Mixture of MotionPrimitives (MoMP) architecture for sign language animation . A set of distinctmotion primitives are learnt during training, that can be temporally combined to animate continuous sign language sequences .…