Convolutional neural networks (CNNs) often have poor generalization performance under domain shift. One way to improve domain generalization is to collect diverse source data from multiple relevant domains so that a CNN model is allowed to learn more
domain-invariant, and hence generalizable representations. In this work, we address domain generalization with MixStyle, a plug-and-play, parameter-free module that is simply inserted to shallow CNN layers and requires no modification to training objectives. Specifically, MixStyle probabilistically mixes feature statistics between instances. This idea is inspired by the observation that visual domains can often be characterized by image styles which are in turn encapsulated within instance-level feature statistics in shallow CNN layers. Therefore, inserting MixStyle modules in effect synthesizes novel domains albeit in an implicit way. MixStyle is not only simple and flexible, but also versatile -- it can be used for problems whereby unlabeled images are available, such as semi-supervised domain generalization and unsupervised domain adaptation, with a simple extension to mix feature statistics between labeled and pseudo-labeled instances. We demonstrate through extensive experiments that MixStyle can significantly boost the out-of-distribution generalization performance across a wide range of tasks including object recognition, instance retrieval, and reinforcement learning.
Gradient-based meta-learning and hyperparameter optimization have seen significant progress recently, enabling practical end-to-end training of neural networks together with many hyperparameters. Nevertheless, existing approaches are relatively expen
sive as they need to compute second-order derivatives and store a longer computational graph. This cost prevents scaling them to larger network architectures. We present EvoGrad, a new approach to meta-learning that draws upon evolutionary techniques to more efficiently compute hypergradients. EvoGrad estimates hypergradient with respect to hyperparameters without calculating second-order gradients, or storing a longer computational graph, leading to significant improvements in efficiency. We evaluate EvoGrad on two substantial recent meta-learning applications, namely cross-domain few-shot learning with feature-wise transformations and noisy label learning with MetaWeightNet. The results show that EvoGrad significantly improves efficiency and enables scaling meta-learning to bigger CNN architectures such as from ResNet18 to ResNet34.
The breakthrough of contrastive learning (CL) has fueled the recent success of self-supervised learning (SSL) in high-level vision tasks on RGB images. However, CL is still ill-defined for low-level vision tasks, such as joint demosaicking and denois
ing (JDD), in the RAW domain. To bridge this methodological gap, we present a novel CL approach on RAW images, residual contrastive learning (RCL), which aims to learn meaningful representations for JDD. Our work is built on the assumption that noise contained in each RAW image is signal-dependent, thus two crops from the same RAW image should have more similar noise distribution than two crops from different RAW images. We use residuals as a discriminative feature and the earth movers distance to measure the distribution divergence for the contrastive loss. To evaluate the proposed CL strategy, we simulate a series of unsupervised JDD experiments with large-scale data corrupted by synthetic signal-dependent noise, where we set a new benchmark for unsupervised JDD tasks with unknown (random) noise variance. Our empirical study not only validates that CL can be applied on distributions (c.f. features), but also exposes the lack of robustness of previous non-ML and SSL JDD methods when the statistics of the noise are unknown, thus providing some further insight into signal-dependent noise problems.
Calibration of neural networks is a topical problem that is becoming increasingly important for real-world use of neural networks. The problem is especially noticeable when using modern neural networks, for which there is significant difference betwe
en the model confidence and the confidence it should have. Various strategies have been successfully proposed, yet there is more space for improvements. We propose a novel approach that introduces a differentiable metric for expected calibration error and successfully uses it as an objective for meta-learning, achieving competitive results with state-of-the-art approaches. Our approach presents a new direction of using meta-learning to directly optimize model calibration, which we believe will inspire further work in this promising and new direction.
Current supervised sketch-based image retrieval (SBIR) methods achieve excellent performance. However, the cost of data collection and labeling imposes an intractable barrier to practical deployment of real applications. In this paper, we present the
first attempt at unsupervised SBIR to remove the labeling cost (category annotations and sketch-photo pairings) that is conventionally needed for training. Existing single-domain unsupervised representation learning methods perform poorly in this application, due to the unique cross-domain (sketch and photo) nature of the problem. We therefore introduce a novel framework that simultaneously performs unsupervised representation learning and sketch-photo domain alignment. Technically this is underpinned by exploiting joint distribution optimal transport (JDOT) to align data from different domains during representation learning, which we extend with trainable cluster prototypes and feature memory banks to further improve scalability and efficacy. Extensive experiments show that our framework achieves excellent performance in the new unsupervised setting, and performs comparably or better than state-of-the-art in the zero-shot setting.
Though convolutional neural networks (CNNs) have demonstrated remarkable ability in learning discriminative features, they often generalize poorly to unseen domains. Domain generalization aims to address this problem by learning from a set of source
domains a model that is generalizable to any unseen domain. In this paper, a novel approach is proposed based on probabilistically mixing instance-level feature statistics of training samples across source domains. Our method, termed MixStyle, is motivated by the observation that visual domain is closely related to image style (e.g., photo vs.~sketch images). Such style information is captured by the bottom layers of a CNN where our proposed style-mixing takes place. Mixing styles of training instances results in novel domains being synthesized implicitly, which increase the domain diversity of the source domains, and hence the generalizability of the trained model. MixStyle fits into mini-batch training perfectly and is extremely easy to implement. The effectiveness of MixStyle is demonstrated on a wide range of tasks including category classification, instance retrieval and reinforcement learning.
Sketch-based image retrieval (SBIR) is a cross-modal matching problem which is typically solved by learning a joint embedding space where the semantic content shared between photo and sketch modalities are preserved. However, a fundamental challenge
in SBIR has been largely ignored so far, that is, sketches are drawn by humans and considerable style variations exist amongst different users. An effective SBIR model needs to explicitly account for this style diversity, crucially, to generalise to unseen user styles. To this end, a novel style-agnostic SBIR model is proposed. Different from existing models, a cross-modal variational autoencoder (VAE) is employed to explicitly disentangle each sketch into a semantic content part shared with the corresponding photo, and a style part unique to the sketcher. Importantly, to make our model dynamically adaptable to any unseen user styles, we propose to meta-train our cross-modal VAE by adding two style-adaptive components: a set of feature transformation layers to its encoder and a regulariser to the disentangled semantic content latent code. With this meta-learning framework, our model can not only disentangle the cross-modal shared semantic content for SBIR, but can adapt the disentanglement to any unseen user style as well, making the SBIR model truly style-agnostic. Extensive experiments show that our style-agnostic model yields state-of-the-art performance for both category-level and instance-level SBIR.
Analysis of human sketches in deep learning has advanced immensely through the use of waypoint-sequences rather than raster-graphic representations. We further aim to model sketches as a sequence of low-dimensional parametric curves. To this end, we
propose an inverse graphics framework capable of approximating a raster or waypoint based stroke encoded as a point-cloud with a variable-degree Bezier curve. Building on this module, we present Cloud2Curve, a generative model for scalable high-resolution vector sketches that can be trained end-to-end using point-cloud data alone. As a consequence, our model is also capable of deterministic vectorization which can map novel raster or waypoint based sketches to their corresponding high-resolution scalable Bezier equivalent. We evaluate the generation and vectorization capabilities of our model on Quick, Draw! and K-MNIST datasets.
We present a novel Tensor Composition Network (TCN) to predict visual relationships in images. Visual Relationships in subject-predicate-object form provide a more powerful query modality than simple image tags. However Visual Relationship Prediction
(VRP) also provides a more challenging test of image understanding than conventional image tagging, and is difficult to learn due to a large label-space and incomplete annotation. The key idea of our TCN is to exploit the low rank property of the visual relationship tensor, so as to leverage correlations within and across objects and relationships, and make a structured prediction of all objects and their relations in an image. To show the effectiveness of our method, we first empirically compare our model with multi-label classification alternatives on VRP, and show that our model outperforms state-of-the-art MLIC methods. We then show that, thanks to our tensor (de)composition layer, our model can predict visual relationships which have not been seen in training dataset. We finally show our TCNs image-level visual relationship prediction provides a simple and efficient mechanism for relation-based image retrieval.
This paper focuses on domain generalization (DG), the task of learning from multiple source domains a model that generalizes well to unseen domains. A main challenge for DG is that the available source domains often exhibit limited diversity, hamperi
ng the models ability to learn to generalize. We therefore employ a data generator to synthesize data from pseudo-novel domains to augment the source domains. This explicitly increases the diversity of available training domains and leads to a more generalizable model. To train the generator, we model the distribution divergence between source and synthesized pseudo-novel domains using optimal transport, and maximize the divergence. To ensure that semantics are preserved in the synthesized data, we further impose cycle-consistency and classification losses on the generator. Our method, L2A-OT (Learning to Augment by Optimal Transport) outperforms current state-of-the-art DG methods on four benchmark datasets.
