WACV 2021 Papers

In this work, we propose an AI-based method that intends to improve the conventional retinal disease treatment procedure and help ophthalmologists increase diagnosis efficiency and accuracy. The proposed method is composed of a deep neural networks-based (DNN-based) module, including a retinal disease identifier and clinical description generator, and a DNN visual explanation module. To train and validate the effectiveness of our DNN-based module, we propose a large-scale retinal disease image dataset. Also, as ground truth, we provide a retinal image dataset manually labeled by ophthalmologists to qualitatively show the proposed AI-based method is effective. With our experimental results, we show that the proposed method is quantitatively and qualitatively effective. Our method is capable of creating meaningful retinal image descriptions and visual explanations that are clinically relevant.

Clothing recognition is the most fundamental AI application challenge within the fashion domain. While existing solutions offer decent recognition accuracy, they are generally slow and require significant computational resources. In this paper we propose a single-stage approach to overcome this obstacle and deliver rapid clothing detection and keypoint estimation. Our solution is based on a multi-target network CenterNet, and we introduce several powerful post-processing techniques to enhance performance. Our most accurate model achieves results comparable to state-of-the-art solutions on the DeepFashion2 dataset, and our light and fast model runs at 17 FPS on the Huawei P40 Pro smartphone. In addition, we achieved second place in the DeepFashion2 Landmark Estimation Challenge 2020 with 0.582 mAP on the test dataset.

The study of neurodegenerative diseases relies on the reconstruction and analysis of the brain cortex from magnetic resonance imaging (MRI). Traditional frameworks for this task like FreeSurfer demand lengthy runtimes, while its accelerated variant FastSurfer still relies on a voxel-wise segmentation which is limited by its resolution to capture narrow continuous objects as cortical surfaces. Having these limitations in mind, we propose DeepCSR, a 3D deep learning framework for cortical surface reconstruction from MRI. Towards this end, we train a neural network model with hypercolumn features to predict implicit surface representations for points in a brain template space. After training, the cortical surface at a desired level of detail is obtained by evaluating surface representations at specific coordinates, and subsequently applying a topology correction algorithm and an isosurface extraction method. Thanks to the continuous nature of this approach and the efficacy of its hypercolumn features scheme, DeepCSR efficiently reconstructs cortical surfaces at high resolution capturing fine details in the cortical folding. Moreover, DeepCSR is as accurate, more precise, and faster than the widely used FreeSurfer toolbox and its deep learning powered variant FastSurfer on reconstructing cortical surfaces from MRI which should facilitate large-scale medical studies and new healthcare applications.

Recently, there is a vast interest in developing image feature learning methods that are independent of the training data, such as deep image prior, InGAN, SinGAN, and DCIL. These methods are unsupervised and are used to perform low-level vision tasks such as image restoration, image editing, and image synthesis. In this work, we proposed a new training data-independent framework, called Deep Contextual Features Learning (DeepCFL), to perform image synthesis and image restoration based on the semantics of the input image. The contextual features are simply the high dimensional vectors representing the semantics of the given image. DeepCFL is a single image GAN framework that learns the distribution of the context vectors from the input image. We show the performance of contextual learning in various challenging scenarios: outpainting, inpainting, and restoration of randomly removed pixels. DeepCFL is applicable when the input source image and the generated target image are not aligned. We illustrate image synthesis using DeepCFL for the task of image resizing.

This paper proposes a novel method to detect anomalies in large datasets under a fully unsupervised setting. The key idea behind our algorithm is to learn the representation underlying normal data. To this end, we leverage the latest clustering technique suitable for handling high dimensional data. This hypothesis provides a reliable starting point for normal data selection. We train an autoencoder from the normal data subset, and iterate between hypothesizing nor-mal candidate subset based on clustering and representation learning. The reconstruction error from the learned autoencoder serves as a scoring function to assess the normality of the data. Experimental results on several public benchmark datasets show that the proposed method outperforms state-of-the-art unsupervised techniques and is comparable to semi-supervised techniques in most cases

Much of the focus in the object detection literature has been on the problem of identifying the bounding box of a particular class of object in an image. Yet, in contexts such as robotics and augmented reality, it is often necessary to find a specific object instance--a unique toy or a custom industrial part for example--rather than a generic object class. Here, applications can require a rapid shift from one object instance to another, thus requiring fast turnaround which affords little-to-no training time. What is more, gathering a dataset and training a model for every new object instance to be detected can be an expensive and time-consuming process. In this context, we propose a generic 2D object instance detection approach that uses example viewpoints of the target object at test time to retrieve its 2D location in RGB images, without requiring any additional training (i.e. fine-tuning) step. To this end, we present an end-to-end architecture that extracts global and local information of the object from its viewpoints. The global information is used to tune early filters in the backbone while local viewpoints are correlated with the input image. Our method offers an improvement of almost 30 mAP over the previous template matching methods on the challenging Occluded Linemod [3] dataset (overall mAP of 50.7). Our experiments also show that our single generic model (not trained on any of the test objects) yields detection results that are on par with approaches that are trained specifically on the target objects.

End-users, without knowledge in photography, desire to beautify their photos to have a similar color style as a well-retouched reference. However, recent works in image style transfer are overused. They usually synthesize undesirable results due to transferring exact colors to the wrong destination. It becomes even worse in sensitive cases such as portraits. In this work, we concentrate on learning low-level image transformation, especially color-shifting methods, rather than contextual features matching, then present a novel supervised approach for color style transfer. Furthermore, we propose a color style transfer named Deep Preset designed to 1) generalize the features representing the color transformation from content with natural colors to retouched reference, then blend it into the contextual features of content, 2) predict hyper-parameters (settings or preset) of the applied low-level color transformation methods, 3) stylize content image to have a similar color style as reference. We script Lightroom, a powerful tool in editing photos, to generate 600,000 training samples using 1,200 images from the Flick2K dataset and 500 user-generated presets with 69 settings. Experimental results show that our Deep Preset outperforms the previous works in color style transfer quantitatively and qualitatively. Our work is available at https://minhmanho.github.io/deep_preset/.

Due to respectively limited training data, different entities addressing the same vision task based on certain sensitive images may not train a robust deep network. This paper introduces a new vision task where various entities share task-specific image data to enlarge each other's training data volume without visually disclosing sensitive contents (e.g. illegal images). Then, we present a new structure-based training regime to enable different entities learn task-specific and reconstruction-proof image representations for image data sharing. Specifically, each entity learns a private Deep Poisoning Module (DPM) and insert it to a pre-trained deep network, which is designed to perform the specific vision task. The DPM deliberately poisons convolutional image features to prevent image reconstructions, while ensuring that the altered image data is functionally equivalent to the non-poisoned data for the specific vision task. Given this equivalence, the poisoned features shared from one entity could be used by another entity for further model refinement. Experimental results on image classification prove the efficacy of the proposed method.

Existing deep learning based interactive segmentation methods have achieved remarkable performance with only a few user clicks, e.g. DEXTR attaining 91.5% IoU on PASCAL VOC with only four extreme clicks. However, we observe even the state-of-the-art methods would often struggle in cases of objects to be segmented with elongated thin structures (e.g. bug legs and bicycle spokes). We investigate such failures, and find the critical reasons behind are two-fold: 1) lack of appropriate training dataset; and 2) extremely imbalanced distribution w.r.t. number of pixels belonging to thin and non-thin regions. Targeted at these challenges, we collect a large-scale dataset specifically for segmentation of thin elongated objects, named ThinObject-5K. Also, we present a novel integrative thin object segmentation network consisting of three streams. Among them, the high-resolution edge stream aims at preserving fine-grained details including elongated thin parts; the fixed-resolution context stream focuses on capturing semantic contexts. The two streams' outputs are then amalgamated in the fusion stream to complement each other for help producing a refined segmentation output with sharper predictions around thin parts. Extensive experimental results well demonstrate the effectiveness of our proposed solution on segmenting thin objects, surpassing the baseline by 30% IoU_thin despite using only four clicks. Codes and dataset are available at https://github.com/liewjunhao/thin-object-selection.

Image compositing is a task of combining regions from different images to compose a new image. A common use case is background replacement of portrait images. To obtain high quality composites, professionals typically manually perform multiple editing steps such as segmentation, matting and foreground color decontamination, which is very time consuming even with sophisticated photo editing tools. In this paper, we propose a new method which can automatically generate high-quality image compositing without any user input. Our method can be trained end-to-end to optimize exploitation of contextual and color information of both foreground and background images, where the compositing quality is considered in the optimization. Specifically, inspired by Laplacian pyramid blending, a denseconnected multi-stream fusion network is proposed to effectively fuse the information from the foreground and background images at different scales. In addition, we introduce a self-taught strategy to progressively train from easy to complex cases to mitigate the lack of training data. Experiments show that the proposed method can automatically generate high-quality composites and outperforms existing methods both qualitatively and quantitatively.

CNN visualization and interpretation methods, like class-activation maps (CAMs), are typically used to highlight the image regions linked to class predictions. These models allow to simultaneously classify images and extract class-dependent saliency maps, without the need for costly pixel-level annotations. However, they typically yield segmentations with high false-positive rates and, therefore, coarse visualisations, more so when processing challenging images, as encountered in histology. To mitigate this issue, we propose an active learning (AL) framework, which progressively integrates pixel-level annotations during training. Given training data with global image-level labels, our deep weakly-supervised learning model jointly performs supervised image-level classification and active learning for segmentation, integrating pixel annotations by an oracle. Unlike standard AL methods that focus on sample selection, we also leverage large numbers of unlabeled images via pseudo-segmentations (i.e., self-learning at the pixel level), and integrate them with the oracle-annotated samples during training. We report extensive experiments over two challenging benchmarks -- high-resolution medical images (histology GlaS data for colon cancer) and natural images (CUB-200-2011 for bird species). Our results indicate that, by simply using random sample selection, the proposed approach can significantly outperform state-of the-art CAMs and AL methods, with an identical oracle-supervision budget. Our code is publicly available.

The imbalanced distribution of the training data makes the networks biased to the frequent classes. Existing methods to resolve the problem involve re-sampling, re-weighting, or cost-sensitive learning. Most of them anticipate that emphasizing the minority classes during the training would help the network to learn better representations. In this paper, we propose a method for reparameterizing softmax classifiers' offsets so that training is less sensitive to class imbalance. We first observe that the trained offset of the baseline linear classifier is biased toward the majority classes due to the imbalance. Instead of the trained offset, we define the estimated offset, and constrain it to be uniform over the classes. In experiments with long-tailed benchmarks, our method exhibits the best performance. These experiments verify that our proposed method effectively encourages the networks to learn better representations for minority classes while preserving the performance for the majority classes.

We present DeepInversion for Object Detection (DIODE) to enable data-free knowledge distillation for neural networks trained on the object detection task. From a data-free perspective, DIODE synthesizes images given only an off-the-shelf pre-trained detection network and without any prior domain knowledge, generator network, or pre-computed activations. DIODE relies on two key components--first, an extensive set of differentiable augmentations to improve image fidelity and distillation effectiveness. Second, a novel automated bounding box and category sampling scheme for image synthesis enabling generating a large number of images with a diverse set of spatial and category objects. The resulting images enable data-free knowledge distillation from a teacher to a student detector, initialized from scratch. In an extensive set of experiments, we demonstrate that DIODE's ability to match the original training distribution consistently enables more effective knowledge distillation than out-of-distribution proxy datasets, which unavoidably occur in a data-free setup given the absence of the original domain knowledge.

The task of video and text sequence alignment is a prerequisite step toward joint understanding of movie videos and screenplays. However, supervised methods face the obstacle of limited realistic training data. With this paper, we attempt to enhance data efficiency of the end-to-end alignment network NeuMATCH [15]. Recent research [56] suggests that network components dealing with different modalities may overfit and generalize at different speeds, creating difficulties for training. We propose to employ (1) layer-wise adaptive rate scaling (LARS) to align the magnitudes of gradient updates in different layers and balance the pace of learning and (2) sequence-wise batch normalization (SBN) to align the internal feature distributions from different modalities. Finally, we leverage random projection to reduce the dimensionality of input features. On the YouTube Movie Summary dataset, the combined use of these technique closes the performance gap when the pretraining on the LSMDC dataset is omitted and achieves the state-of-the-art result. Extensive empirical comparisons and analysis reveal that these techniques improve optimization and regularize the network more effectively than two different setups of layer normalization.

This paper studies the task of temporal moment localization in a long untrimmed video using natural language query. Given a query sentence, the goal is to determine the start and end of the relevant segment within the video. Our key innovation is to learn a video feature embedding through a language-conditioned message-passing algorithm suitable for temporal moment localization which captures the relationships between humans, objects and activities in the video. These relationships are obtained by a spatial subgraph that contextualized the scene representation using detected objects and human features. Moreover, a temporal sub-graph captures the activities within the video through time. Our method is evaluated on three standard benchmark datasets, and we also introduce YouCookII as a new benchmark for this task. Experiments show our method outperforms state-of-the-art methods on these datasets, confirming the effectiveness of our approach

Driven by deep learning, object recognition has recently made a tremendous leap forward. Nonetheless, its accuracy often still suffers from several sources of variation that can be found in real-world images. Some of the most challenging variation are induced by changing lighting conditions. This paper presents a novel approach for tackling bright-ness variation in the domain of 2D object detection and 6D object pose estimation. Existing works aiming at improving robustness towards different lighting conditions are of-ten grounded on classical computer vision contrast normalisation techniques or the acquisition of large amounts of an-notated data in order to achieve invariance during training.While the former cannot generalise well to a wide range of illumination conditions, the latter is neither practical nor scalable. Hence, we propose the usage of Generative Adversarial Network in order to learn how to normalise the illumination of an input image. Thereby, the generator is explicitly designed to normalise illumination in images soto enhance the object recognition performance. Extensive evaluations demonstrate that leveraging the generated data can significantly enhance the detection performance, out-performing all other state-of-the-art methods. We further constitute a natural extension focusing on white balance variations and introduce a new dataset for evaluation.

Contour-based instance segmentation methods are attractive due to their efficiency. However, existing contour-based methods either suffer from lossy representation, complex pipeline or difficulty in model training, resulting in subpar mask accuracy on challenging datasets like MS-COCO. In this work, we propose a novel deep attentive contour model, named DANCE, to achieve better instance segmentation accuracy while remaining good efficiency. To this end, DANCE applies two new designs: attentive contour deformation to refine the quality of segmentation contours and segment-wise matching to ease the model training. Comprehensive experiments demonstrate DANCE excels at deforming the initial contour in a more natural and efficient way towards the real object boundaries. Effectiveness of DANCE is also validated on the COCO dataset, which achieves 38.1% mAP and outperforms all other contour-based instance segmentation models. To the best of our knowledge, DANCE is the first contour-based model that achieves comparable performance to pixel-wise segmentation models. Code is available at https://github.com/lkevinzc/dance.

Semantic segmentation models based on convolutional neural networks have recently displayed remarkable performance for a multitude of applications. However, these models typically do not generalize well when applied on new domains, especially when going from synthetic to real data. In this paper we address the problem of unsupervised domain adaptation (UDA), which attempts to train on labelled data from one domain (source domain), and simultaneously learn from unlabelled data in the domain of interest (target domain). Existing methods have seen success by training on pseudo-labels for these unlabelled images. Multiple techniques have been proposed to mitigate low-quality pseudo-labels arising from the domain shift, with varying degrees of success. We propose DACS: Domain Adaptation via Cross-domain mixed Sampling, which mixes images from the two domains along with the corresponding labels and pseudo-labels. These mixed samples are then trained on, in addition to the labelled data itself. We demonstrate the effectiveness of our solution by achieving state-of-the-art results for GTA5 to Cityscapes, a common synthetic-to-real semantic segmentation benchmark for UDA.

In this paper, we focus on exploring the fusion of images and point clouds for 3D object detection in view of the complementary nature of the two modalities, i.e., images possess more semantic information while point clouds specialize in distance sensing. To this end, we present a novel two-stage multi-modal fusion network for 3D object detection, taking both binocular images and raw point clouds as input. The whole architecture facilitates two-stage fusion. The first stage aims at producing 3D proposals through point-wise feature fusion. Within the first stage, we further exploit a joint anchor mechanism that enables the network to utilize 2D-3D classification and regression simultaneously for better proposal generation. The second stage works on the 2D and 3D proposal regions and fuses their features. In addition, we propose to use pseudo LiDAR points from stereo matching as a data augmentation method to densify the LiDAR points, as we observe that objects missed by the detection network mostly have too few points, especially for far-away objects. Our experiments on the KITTI dataset show that the proposed multi-stage fusion helps the network to learn better representations.