Abstract: Top-performing computer vision models are powered by convolutional neural networks (CNNs). Training an accurate CNN highly depends on both the raw sensor data and their associated ground truth (GT). Collecting such GT is usually done through human labeling, which is time-consuming and does not scale as we wish. This data-labeling bottleneck may be intensified due to domain shifts among image sensors, which could force per-sensor data labeling. In this paper, we focus on the use of co-training, a semi-supervised learning (SSL) method, for obtaining self-labeled object bounding boxes (BBs), i.e., the GT to train deep object detectors. In particular, we assess the goodness of multi-modal co-training by relying on two different views of an image, namely, appearance (RGB) and estimated depth (D). Moreover, we compare appearance-based single-modal co-training with multi-modal. Our results suggest that in a standard SSL setting (no domain shift, a few human-labeled data) and under virtual-to-real domain shift (many virtual-world labeled data, no human-labeled data) multi-modal co-training outperforms single-modal. In the latter case, by performing GAN-based domain translation both co-training modalities are on par, at least when using an off-the-shelf depth estimation model not specifically trained on the translated images.

Abstract: BACKGROUND : Artificial intelligence (AI) research in colonoscopy is progressing rapidly but widespread clinical implementation is not yet a reality. We aimed to identify the top implementation research priorities. METHODS : An established modified Delphi approach for research priority setting was used. Fifteen international experts, including endoscopists and translational computer scientists/engineers, from nine countries participated in an online survey over 9 months. Questions related to AI implementation in colonoscopy were generated as a long-list in the first round, and then scored in two subsequent rounds to identify the top 10 research questions. RESULTS : The top 10 ranked questions were categorized into five themes. Theme 1: clinical trial design/end points (4 questions), related to optimum trial designs for polyp detection and characterization, determining the optimal end points for evaluation of AI, and demonstrating impact on interval cancer rates. Theme 2: technological developments (3 questions), including improving detection of more challenging and advanced lesions, reduction of false-positive rates, and minimizing latency. Theme 3: clinical adoption/integration (1 question), concerning the effective combination of detection and characterization into one workflow. Theme 4: data access/annotation (1 question), concerning more efficient or automated data annotation methods to reduce the burden on human experts. Theme 5: regulatory approval (1 question), related to making regulatory approval processes more efficient. CONCLUSIONS : This is the first reported international research priority setting exercise for AI in colonoscopy. The study findings should be used as a framework to guide future research with key stakeholders to accelerate the clinical implementation of AI in endoscopy.

Abstract: Weather prediction from real-world images can be termed a complex task when targeting classification using neural networks. Moreover, the number of images throughout the available datasets can contain a huge amount of variance when comparing locations with the weather those images are representing. In this article, the capabilities of a custom built driver simulator are explored specifically to simulate a wide range of weather conditions. Moreover, the performance of a new synthetic dataset generated by the above simulator is also assessed. The results indicate that the use of synthetic datasets in conjunction with real-world datasets can increase the training efficiency of the CNNs by as much as 74%. The article paves a way forward to tackle the persistent problem of bias in vision-based datasets.

Abstract: Aesthetic image analysis is essential for improving the performance of multimedia image retrieval systems, especially from a repository of social media and multimedia content stored on mobile devices. This paper presents a novel method for classifying aesthetic natural scene images by studying the naturalness of image content using statistical features, and reading text in the images using semantic features. Unlike existing methods that focus only on image quality with human information, the proposed approach focuses on image features as well as text-based semantic features without human intervention to reduce the gap between subjectivity and objectivity in the classification. The aesthetic classes considered in this work are (i) Very Pleasant, (ii) Pleasant, (iii) Normal and (iv) Unpleasant. The naturalness is represented by features of focus, defocus, perceived brightness, perceived contrast, blurriness and noisiness, while semantics are represented by text recognition, description of the images and labels of images, profile pictures, and banner images. Furthermore, a deep learning model is proposed in a novel way to fuse statistical and semantic features for the classification of aesthetic natural scene images. Experiments on our own dataset and the standard datasets demonstrate that the proposed approach achieves 92.74%, 88.67% and 83.22% average classification rates on our own dataset, AVA dataset and CUHKPQ dataset, respectively. Furthermore, a comparative study of the proposed model with the existing methods shows that the proposed method is effective for the classification of aesthetic social media images.

Abstract: A long-standing view is that lightness illusions are by-products of strategies employed by the visual system to stabilize its perceptual representation of surface reflectance against changes in illumination. Computationally, one such strategy is to infer reflectance from the retinal image, and to base the lightness percept on this inference. CNNs trained to infer reflectance from images have proven successful at solving this problem under limited conditions. To evaluate whether these CNNs provide suitable starting points for computational models of human lightness perception, we tested a state-of-the-art CNN on several lightness illusions, and compared its behaviour to prior measurements of human performance. We trained a CNN (Yu & Smith, 2019) to infer reflectance from luminance images. The network had a 30-layer hourglass architecture with skip connections. We trained the network via supervised learning on 100K images, rendered in Blender, each showing randomly placed geometric objects (surfaces, cubes, tori, etc.), with random Lambertian reflectance patterns (solid, Voronoi, or low-pass noise), under randomized point+ambient lighting. The renderer also provided the ground-truth reflectance images required for training. After training, we applied the network to several visual illusions. These included the argyle, Koffka-Adelson, snake, White’s, checkerboard assimilation, and simultaneous contrast illusions, along with their controls where appropriate. The CNN correctly predicted larger illusions in the argyle, Koffka-Adelson, and snake images than in their controls. It also correctly predicted an assimilation effect in White's illusion. It did not, however, account for the checkerboard assimilation or simultaneous contrast effects. These results are consistent with the view that at least some lightness phenomena are by-products of a rational approach to inferring stable representations of physical properties from intrinsically ambiguous retinal images. Furthermore, they suggest that CNN models may be a promising starting point for new models of human lightness perception.

Abstract: This paper presents and evaluates a framework to improve the performance of visual object classification methods, which are based on the usage of image feature descriptors as inputs. The goal of the proposed framework is to learn the best descriptor for each image in a given database. This goal is reached by means of a reinforcement learning process using the minimum information. The visual classification system used to demonstrate the proposed framework is based on a bag of features scheme, and the reinforcement learning technique is implemented through the Q-learning approach. The behavior of the reinforcement learning with different state definitions is evaluated. Additionally, a method that combines all these states is formulated in order to select the optimal state. Finally, the chosen actions are obtained from the best set of image descriptors in the literature: PHOW, SIFT, C-SIFT, SURF and Spin. Experimental results using two public databases (ETH and COIL) are provided showing both the validity of the proposed approach and comparisons with state of the art. In all the cases the best results are obtained with the proposed approach.

Abstract: Human multi-limb segmentation in RGB images has attracted a lot of interest in the research community because of the huge amount of possible applications in fields like Human-Computer Interaction, Surveillance, eHealth, or Gaming. Nevertheless, human multi-limb segmentation is a very hard task because of the changes in appearance produced by different points of view, clothing, lighting conditions, occlusions, and number of articulations of the human body. Furthermore, this huge pose variability makes the availability of large annotated datasets difficult. In this paper, we introduce the HuPBA8k+ dataset. The dataset contains more than 8000 labeled frames at pixel precision, including more than 120000 manually labeled samples of 14 different limbs. For completeness, the dataset is also labeled at frame-level with action annotations drawn from an 11 action dictionary which includes both single person actions and person-person interactive actions. Furthermore, we also propose a two-stage approach for the segmentation of human limbs. In a first stage, human limbs are trained using cascades of classifiers to be split in a tree-structure way, which is included in an Error-Correcting Output Codes (ECOC) framework to define a body-like probability map. This map is used to obtain a binary mask of the subject by means of GMM color modelling and GraphCuts theory. In a second stage, we embed a similar tree-structure in an ECOC framework to build a more accurate set of limb-like probability maps within the segmented user mask, that are fed to a multi-label GraphCut procedure to obtain final multi-limb segmentation. The methodology is tested on the novel HuPBA8k+ dataset, showing performance improvements in comparison to state-of-the-art approaches. In addition, a baseline of standard action recognition methods for the 11 actions categories of the novel dataset is also provided.

Abstract: In this paper, we propose a novel technique to detect the shadows cast by urban infrastructure, such as buildings, billboards, and traffic signs, using a sequence of images taken from a fixed camera. In our approach, we compute two different background models in parallel: one for the edges and one for the reflected light intensity. An algorithm is proposed to train the system to distinguish between moving edges in general and edges that belong to static objects, creating an edge background model. Then, during operation, a background intensity model allow us to separate between moving and static objects. Those edges included in the moving objects and those that belong to the edge background model are subtracted from the current image edges. The remaining edges are the ones cast by urban infrastructure. Our method is tested on a typical crossroad scene and the results show that the approach is sound and promising.