Attention modules connecting encoder and decoders have been widely applied in the field of object recognition, image captioning, visual question answering and neural machine translation, and significantly improves the performance. In this paper, we propose a bottom-up gated hierarchical attention (GHA) mechanism for image captioning. Our proposed model employs a CNN as the decoder which is able to learn different concepts at different layers, and apparently, different concepts correspond to different areas of an image. Therefore, we develop the GHA in which low-level concepts are merged into high-level concepts and simultaneously low-level attended features pass to the top to make predictions. Our GHA significantly improves the performance of the model that only applies one level attention, for example, the CIDEr score increases from 0.923 to 0.999, which is comparable to the state-of-the-art models that employ attributes boosting and reinforcement learning (RL). We also conduct extensive experiments to analyze the CNN decoder and our proposed GHA, and we find that deeper decoders cannot obtain better performance, and when the convolutional decoder becomes deeper the model is likely to collapse during training.
We propose a deep learning clustering method that exploits dense features from a segmentation network for emphysema subtyping from computed tomography (CT) scans. Using dense features enables high-resolution visualization of image regions corresponding to the cluster assignment via dense clustering activation maps (dCAMs). This approach provides model interpretability. We evaluated clustering results on 500 subjects from the COPDGenestudy, where radiologists manually annotated emphysema sub-types according to their visual CT assessment. We achieved a 43% unsupervised clustering accuracy, outperforming our baseline at 41% and yielding results comparable to supervised classification at 45%. The proposed method also offers a better cluster formation than the baseline, achieving0.54 in silhouette coefficient and 0.55 in David-Bouldin scores.
Most existing face image Super-Resolution (SR) methods assume that the Low-Resolution (LR) images were artificially downsampled from High-Resolution (HR) images with bicubic interpolation. This operation changes the natural image characteristics and reduces noise. Hence, SR methods trained on such data most often fail to produce good results when applied to real LR images. To solve this problem, we propose a novel framework for generation of realistic LR/HR training pairs. Our framework estimates realistic blur kernels, noise distributions, and JPEG compression artifacts to generate LR images with similar image characteristics as the ones in the source domain. This allows us to train a SR model using high quality face images as Ground-Truth (GT). For better perceptual quality we use a Generative Adversarial Network (GAN) based SR model where we have exchanged the commonly used VGG-loss [24] with LPIPS-loss [52]. Experimental results on both real and artificially corrupted face images show that our method results in more detailed reconstructions with less noise compared to existing State-of-the-Art (SoTA) methods. In addition, we show that the traditional non-reference Image Quality Assessment (IQA) methods fail to capture this improvement and demonstrate that the more recent NIMA metric [16] correlates better with human perception via Mean Opinion Rank (MOR).
Object localization has a vital role in any object detector, and therefore, has been the focus of attention by many researchers. In this article, a special training approach is proposed for a light convolutional neural network (CNN) to determine the region of interest (ROI) in an image while effectively reducing the number of probable anchor boxes. Almost all CNN-based detectors utilize a fixed input size image, which may yield poor performance when dealing with various object sizes. In this paper, a different CNN structure is proposed taking three different input sizes, to enhance the performance. In order to demonstrate the effectiveness of the proposed method, two common data set are used for training while tracking by localization application is considered to demonstrate its final performance. The promising results indicate the applicability of the presented structure and the training method in practice.
Video frame interpolation is the task of creating an interface between two adjacent frames along the time axis. So, instead of simply averaging two adjacent frames to create an intermediate image, this operation should maintain semantic continuity with the adjacent frames. Most conventional methods use optical flow, and various tools such as occlusion handling and object smoothing are indispensable. Since the use of these various tools leads to complex problems, we tried to tackle the video interframe generation problem without using problematic optical flow. To enable this, we have tried to use a deep neural network with an invertible structure and developed an invertible U-Net which is a modified normalizing flow. In addition, we propose a learning method with a new consistency loss in the latent space to maintain semantic temporal consistency between frames. The resolution of the generated image is guaranteed to be identical to that of the original images by using an invertible network. Furthermore, as it is not a random image like the ones by generative models, our network guarantees stable outputs without flicker. Through experiments, we confirmed the feasibility of the proposed algorithm and would like to suggest invertible U-Net as a new possibility for baseline in video frame interpolation. This paper is meaningful in that it is the worlds first attempt to use invertible networks instead of optical flows for video interpolation.
Most, if not all, modern deep learning systems restrict themselves to a single dataset for neural network training and inference. In this article, we are interested in systematic ways to join datasets that are made of similar purposes. Unlike previous published works that ubiquitously conduct the dataset joining in the uninterpretable latent vectorial space, the core to our method is an augmentation procedure in the label space. The primary challenge to address the label space for dataset joining is the discrepancy between labels: non-overlapping label annotation sets, different labeling granularity or hierarchy and etc. Notably we propose a new technique leveraging artificially created knowledge graph, recurrent neural networks and policy gradient that successfully achieve the dataset joining in the label space. Empirical results on both image and text classification justify the validity of our approach.
Generative Adversarial Networks (GANs) are commonly used for modeling complex distributions of data. Both the generators and discriminators of GANs are often modeled by neural networks, posing a non-transparent optimization problem which is non-convex and non-concave over the generator and discriminator, respectively. Such networks are often heuristically optimized with gradient descent-ascent (GDA), but it is unclear whether the optimization problem contains any saddle points, or whether heuristic methods can find them in practice. In this work, we analyze the training of Wasserstein GANs with two-layer neural network discriminators through the lens of convex duality, and for a variety of generators expose the conditions under which Wasserstein GANs can be solved exactly with convex optimization approaches, or can be represented as convex-concave games. Using this convex duality interpretation, we further demonstrate the impact of different activation functions of the discriminator. Our observations are verified with numerical results demonstrating the power of the convex interpretation, with applications in progressive training of convex architectures corresponding to linear generators and quadratic-activation discriminators for CelebA image generation. The code for our experiments is available at https://github.com/ardasahiner/ProCoGAN.
Given a monocular colour image of a warehouse rack, we aim to predict the bird's-eye view layout for each shelf in the rack, which we term as multi-layer layout prediction. To this end, we present RackLay, a deep neural network for real-time shelf layout estimation from a single image. Unlike previous layout estimation methods, which provide a single layout for the dominant ground plane alone, RackLay estimates the top-view and front-view layout for each shelf in the considered rack populated with objects. RackLay's architecture and its variants are versatile and estimate accurate layouts for diverse scenes characterized by varying number of visible shelves in an image, large range in shelf occupancy factor and varied background clutter. Given the extreme paucity of datasets in this space and the difficulty involved in acquiring real data from warehouses, we additionally release a flexible synthetic dataset generation pipeline WareSynth which allows users to control the generation process and tailor the dataset according to contingent application. The ablations across architectural variants and comparison with strong prior baselines vindicate the efficacy of RackLay as an apt architecture for the novel problem of multi-layered layout estimation. We also show that fusing the top-view and front-view enables 3D reasoning applications such as metric free space estimation for the considered rack.
We address the problem of text-guided video temporal grounding, which aims to identify the time interval of certain event based on a natural language description. Different from most existing methods that only consider RGB images as visual features, we propose a multi-modal framework to extract complementary information from videos. Specifically, we adopt RGB images for appearance, optical flow for motion, and depth maps for image structure. While RGB images provide abundant visual cues of certain event, the performance may be affected by background clutters. Therefore, we use optical flow to focus on large motion and depth maps to infer the scene configuration when the action is related to objects recognizable with their shapes. To integrate the three modalities more effectively and enable inter-modal learning, we design a dynamic fusion scheme with transformers to model the interactions between modalities. Furthermore, we apply intra-modal self-supervised learning to enhance feature representations across videos for each modality, which also facilitates multi-modal learning. We conduct extensive experiments on the Charades-STA and ActivityNet Captions datasets, and show that the proposed method performs favorably against state-of-the-art approaches.
Invariance to a broad array of image corruptions, such as warping, noise, or color shifts, is an important aspect of building robust models in computer vision. Recently, several new data augmentations have been proposed that significantly improve performance on ImageNet-C, a benchmark of such corruptions. However, there is still a lack of basic understanding on the relationship between data augmentations and test-time corruptions. To this end, we develop a feature space for image transforms, and then use a new measure in this space between augmentations and corruptions called the Minimal Sample Distance to demonstrate there is a strong correlation between similarity and performance. We then investigate recent data augmentations and observe a significant degradation in corruption robustness when the test-time corruptions are sampled to be perceptually dissimilar from ImageNet-C in this feature space. Our results suggest that test error can be improved by training on perceptually similar augmentations, and data augmentations may not generalize well beyond the existing benchmark. We hope our results and tools will allow for more robust progress towards improving robustness to image corruptions.