The problem of domain generalization is to learn from multiple training domains, and extract a domain-agnostic model that can then be applied an unseen domain. Domain (DG) has clear motivation in contexts where there are target domains with distinct characteristics, yet sparse data for training. For example recognition sketch images, which distinctly more abstract rarer than photos. Nevertheless, DG methods have primarily been evaluated on photo-only benchmarks focusing alleviating the...

Domain shift refers to the well known problem that a model trained in one source domain performs poorly when appliedto target with different statistics. Generalization (DG) techniques attempt alleviate this issue by producing models which design generalize novel testing domains. We propose meta-learning method for generalization. Rather than designing specific is robust as most previous DG work, we agnostic training procedure DG. Our algorithm simulates train/test during synthesizing virtual...

We investigate the problem of fine-grained sketch-based image retrieval (SBIR), where free-hand human sketches are used as queries to perform instance-level images. This is an extremely challenging task because (i) visual comparisons not only need be but also executed cross-domain, (ii) (finger) highly abstract, making matching harder, and most importantly (iii) annotated cross-domain sketch-photo datasets required for training scarce, many state-of-the-art machine learning techniques. In...

Domain generalization (DG) is the challenging and topical problem of learning models that generalize to novel testing domains with different statistics than a set known training domains. The simple approach aggregating data from all source single deep neural network end-to-end on provides surprisingly strong baseline surpasses many prior published methods. In this paper we build by designing an episodic procedure trains in way exposes it domain shift characterises at runtime. Specifically,...

Key for solving fine-grained image categorization is finding discriminate and local regions that correspond to subtle visual traits. Great strides have been made, with complex networks designed specifically learn part-level feature representations. In this paper, we show it possible cultivate details without the need overly complicated network designs or training mechanisms -- a single loss all takes. The main trick lies how delve into individual channels early on, as opposed convention of...

Human sketches are unique in being able to capture both the spatial topology of a visual object, as well its subtle appearance details. Fine-grained sketch-based image retrieval (FG-SBIR) importantly leverages on such fine-grained characteristics conduct instance-level photos. Nevertheless, human often highly abstract and iconic, resulting severe misalignments with candidate photos which turn make detail matching difficult. Existing FG-SBIR approaches focus only coarse holistic via deep...

Fine-grained image analysis (FGIA) is a longstanding and fundamental problem in computer vision pattern recognition, underpins diverse set of real-world applications. The task FGIA targets analyzing visual objects from subordinate categories, e.g., species birds or models cars. small inter-class large intra-class variation inherent to fine-grained makes it challenging problem. Capitalizing on advances deep learning, recent years we have witnessed remarkable progress learning powered FGIA. In...

Sketch-based image retrieval (SBIR) is a challenging task due to the ambiguity inherent in sketches when compared with photos. In this paper, we propose novel convolutional neural network based on Siamese for SBIR. The main idea pull output feature vectors closer input sketch-image pairs that are labeled as similar, and push them away if irrelevant. This achieved by jointly tuning two networks which linked one loss function. Experimental results Flickr15K demonstrate proposed method offers...

We aim to learn a domain generalizable person re-identification (ReID) model. When such model is trained on set of source domains (ReID datasets collected from different camera networks), it can be directly applied any new unseen dataset for effective ReID without updating. Despite its practical value in real-world deployments, has seldom been studied. In this work, novel deep termed Domain-Invariant Mapping Network (DIMN) proposed. DIMN designed mapping between image and identity...

A common strategy adopted by existing state-of-the-art unsupervised domain adaptation (UDA) methods is to employ two classifiers identify the misaligned local regions between source and target domain. Following 'wisdom of crowd' principle, one has ask: why stop at two? Indeed, we find that using more leads better performance, but also introduces model parameters, therefore risking overfitting. In this paper, introduce a novel method called STochastic clAssifieRs (STAR) for addressing...

A few-shot semantic segmentation model is typically composed of a CNN encoder, decoder and simple classifier (separating foreground background pixels). Most existing methods meta-learn all three components for fast adaptation to new class. However, given that as few single support set image available, effective adaption the class extremely challenging. In this work we propose simplify meta-learning task by focusing solely on simplest component – classifier, whilst leaving en-coder...

Free-hand sketches are highly illustrative, and have been widely used by humans to depict objects or stories from ancient times the present. The recent prevalence of touchscreen devices has made sketch creation a much easier task than ever consequently sketch-oriented applications increasingly popular. progress deep learning immensely benefited free-hand research applications. This paper presents comprehensive survey techniques oriented at data, that they enable. main contents this include:...

Image-based virtual try-on aims to fit an in-shop garment into a clothed person image. To achieve this, key step is warping which spatially aligns the target with corresponding body parts in Prior methods typically adopt local appearance flow estimation model. They are thus intrinsically susceptible difficult poses/occlusions and large mis-alignments between images (see Fig. 1). overcome this limitation, novel global model proposed work. For first time, StyleGAN based architecture adopted...

In this paper, we leverage CLIP for zero-shot sketch based image retrieval (ZS-SBIR). We are largely inspired by recent advances on foundation models and the unparalleled generalisation ability they seem to offer, but first time tailor it benefit community. put forward novel designs how best achieve synergy, both category setting fine-grained ('all”}. At very core of our solution is a prompt learning setup. First show just via factoring in sketch-specific prompts, already have category-level...

We propose a multi-scale multi-channel deep neural network framework that, for the first time, yields sketch recognition performance surpassing that of humans. Our superior is result explicitly embedding unique characteristics sketches in our model: (i) architecture designed rather than natural photo statistics, (ii) generalisation encodes sequential ordering sketching process, and (iii) ensemble with joint Bayesian fusion accounts different levels abstraction exhibited free-hand sketches....

We propose a deep hashing framework for sketch retrieval that, the first time, works on multi-million scale human dataset. Leveraging this large dataset, we explore few sketch-specific traits that were otherwise under-studied in prior literature. Instead of following conventional recognition task, introduce novel problem which is not only more challenging, but also offers better testbed large-scale analysis, since: (i) fine-grained feature learning required to accommodate variations style...

In this paper, we develop a novel variational Bayesian learning method for the Dirichlet process (DP) mixture of inverted distributions, which has been shown to be very flexible modeling vectors with positive elements. The recently proposed extended inference (EVI) framework is adopted derive an analytically tractable solution. convergency algorithm theoretically guaranteed by introducing single lower bound approximation original objective function in EVI framework. principle, model can...

Fine-grained sketch-based image retrieval (FG-SBIR) addresses the problem of retrieving a particular photo instance given user's query sketch. Its widespread applicability is however hindered by fact that drawing sketch takes time, and most people struggle to draw complete faithful In this paper, we reformulate conventional FG-SBIR framework tackle these challenges, with ultimate goal target least number strokes possible. We further propose an on-the-fly design starts as soon user drawing....

To see is to sketch - free-hand sketching naturally builds ties between human and machine vision. In this paper, we present a novel approach for translating an object photo sketch, mimicking the process. This extremely challenging task because domains differ significantly. Furthermore, sketches exhibit various levels of sophistication abstraction even when depicting same instance in reference photo. means that if photo-sketch pairs are available, they only provide weak supervision signal...

We propose a perceptual grouping framework that organizes image edges into meaningful structures and demonstrate its usefulness on various computer vision tasks. Our grouper formulates edge as graph partition problem, where learning to rank method is developed encode probabilities of candidate pairs. In particular, RankSVM employed for the first time combine multiple Gestalt principles cue grouping. Afterwards, an based object proposal measure introduced yields proposals comparable...

Fine-grained sketch-based image retrieval (FG-SBIR) addresses matching specific photo instance using free-hand sketch as a query modality. Existing models aim to learn an embedding space in which and can be directly compared. While successful, they require instance-level pairing within each coarse-grained category annotated training data. Since the learned is domain-specific, these do not generalise well across categories. This limits practical applicability of FG-SBIR. In this paper, we...