Abstract: Image to image translation aims to learn a mapping that transforms an image from one visual domain to another. Recent works assume that images descriptors can be disentangled into a domain-invariant content representation and a domain-specific style representation. Thus, translation models seek to preserve the content of source images while changing the style to a target visual domain. However, synthesizing new images is extremely challenging especially in multi-domain translations, as the network has to compose content and style to generate reliable and diverse images in multiple domains. In this paper we propose the use of an image retrieval system to assist the image-to-image translation task. First, we train an image-to-image translation model to map images to multiple domains. Then, we train an image retrieval model using real and generated images to find images similar to a query one in content but in a different domain. Finally, we exploit the image retrieval system to fine-tune the image-to-image translation model and generate higher quality images. Our experiments show the effectiveness of the proposed solution and highlight the contribution of the retrieval network, which can benefit from additional unlabeled data and help image-to-image translation models in the presence of scarce data.

Abstract: The assumption that the retina’s main objective is to provide a minimum entropy representation to higher visual areas (ie efficient coding principle) allows to predict retinal filtering in space–time and colour (Atick, 1992 Network 3 213–251). This is achieved by considering the power spectra of natural images (which is proportional to 1/f2) and the suppression of retinal and image noise. However, most studies consider images within a limited range of lighting conditions (eg near noon) whereas the visual system’s spatial filtering depends on light intensity and the spatiochromatic properties of natural scenes depend of the time of the day. Here, we explore whether the dependence of visual spatial filtering on luminance match the changes in power spectrum of natural scenes at different times of the day. Using human cone-activation based naturalistic stimuli (from the Barcelona Calibrated Images Database), we show that for a range of luminance levels, the shape of the retinal CSF reflects the slope of the power spectrum at low spatial frequencies. Accordingly, the retina implements the filtering which best decorrelates the input signal at every luminance level. This result is in line with the body of work that places efficient coding as a guiding neural principle.

Abstract: We analyze the results of the 2017 ChaLearn Looking at People Challenge at ICCV. The challenge comprised three tracks: (1) large-scale isolated (2) continuous gesture recognition, and (3) real versus fake expressed emotions tracks. It is the second round for both gesture recognition challenges, which were held first in the context of the ICPR 2016 workshop on “multimedia challenges beyond visual analysis”. In this second round, more participants joined the competitions, and the performances considerably improved compared to the first round. Particularly, the best recognition accuracy of isolated gesture recognition has improved from 56.90% to 67.71% in the IsoGD test set, and Mean Jaccard Index (MJI) of continuous gesture recognition has improved from 0.2869 to 0.6103 in the ConGD test set. The third track is the first challenge on real versus fake expressed emotion classification, including six emotion categories, for which a novel database was introduced. The first place was shared between two teams who achieved 67.70% averaged recognition rate on the test set. The data of the three tracks, the participants' code and method descriptions are publicly available to allow researchers to keep making progress in the field.

Abstract: In our work, we are particularly interested in studying the shadows cast by static objects in outdoor environments, during daytime. To assess the accuracy of a shadow detection algorithm, we need ground truth information. The collection of such information is a very tedious task because it is a process that requires manual annotation. To overcome this severe limitation, we propose in this paper a methodology to automatically render ground truth using a virtual environment. To increase the degree of realism and usefulness of the simulated environment, we incorporate in the scenario the precise longitude, latitude and elevation of the actual location of the object, as well as the sun’s position for a given time and day. To evaluate our method, we consider a qualitative and a quantitative comparison. In the quantitative one, we analyze the shadow cast by a real object in a particular geographical location and its corresponding rendered model. To evaluate qualitatively the methodology, we use some ground truth images obtained both manually and automatically.

Abstract: Video retrieval using natural language queries has attracted increasing interest due to its relevance in real-world applications, from intelligent access in private media galleries to web-scale video search. Learning the cross-similarity of video and text in a joint embedding space is the dominant approach. To do so, a contrastive loss is usually employed because it organizes the embedding space by putting similar items close and dissimilar items far. This framework leads to competitive recall rates, as they solely focus on the rank of the groundtruth items. Yet, assessing the quality of the ranking list is of utmost importance when considering intelligent retrieval systems, since multiple items may share similar semantics, hence a high relevance. Moreover, the aforementioned framework uses a fixed margin to separate similar and dissimilar items, treating all non-groundtruth items as equally irrelevant. In this paper we propose to use a variable margin: we argue that varying the margin used during training based on how much relevant an item is to a given query, i.e. a relevance-based margin, easily improves the quality of the ranking lists measured through nDCG and mAP. We demonstrate the advantages of our technique using different models on EPIC-Kitchens-100 and YouCook2. We show that even if we carefully tuned the fixed margin, our technique (which does not have the margin as a hyper-parameter) would still achieve better performance. Finally, extensive ablation studies and qualitative analysis support the robustness of our approach. Code will be released at \urlhttps://github.com/aranciokov/RelevanceMargin-ICMR22.