skip to main content


Title: Place-centric Visual Urban Perception with Deep Multi-instance Regression
This paper presents a unified framework to learn to quantify perceptual attributes (e.g., safety, attractiveness) of physical urban environments using crowd-sourced street-view photos without human annotations. The efforts of this work include two folds. First, we collect a large-scale urban image dataset in multiple major cities in U.S.A., which consists of multiple street-view photos for every place. Instead of using subjective annotations as in previous works, which are neither accurate nor consistent, we collect for every place the safety score from government’s crime event records as objective safety indicators. Second, we observe that the place-centric perception task is by nature a multi-instance regression problem since the labels are only available for places (bags), rather than images or image regions (instances). We thus introduce a deep convolutional neural network (CNN) to parameterize the instance-level scoring function, and develop an EM algorithm to alternatively estimate the primary instances (images or image regions) which affect the safety scores and train the proposed network. Our method is capable of localizing interesting images and image regions for each place.We evaluate the proposed method on a newly created dataset and a public dataset. Results with comparisons showed that our method can clearly outperform the alternative perception methods and more importantly, is capable of generating region-level safety scores to facilitate interpretations of the perception process.  more » « less
Award ID(s):
1657600
NSF-PAR ID:
10056962
Author(s) / Creator(s):
; ; ; ;
Date Published:
Journal Name:
ACM Conference on Multimedia
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Many algorithms for virtual tree generation exist, but the visual realism of the 3D models is unknown. This problem is usually addressed by performing limited user studies or by a side-by-side visual comparison. We introduce an automated system for realism assessment of the tree model based on their perception. We conducted a user study in which 4,000 participants compared over one million pairs of images to collect subjective perceptual scores of a large dataset of virtual trees. The scores were used to train two neural-network-based predictors. A view independent ICTreeF uses the tree model's geometric features that are easy to extract from any model. The second is ICTreeI that estimates the perceived visual realism of a tree from its image. Moreover, to provide an insight into the problem, we deduce intrinsic attributes and evaluate which features make trees look like real trees. In particular, we show that branching angles, length of branches, and widths are critical for perceived realism. We also provide three datasets: carefully curated 3D tree geometries and tree skeletons with their perceptual scores, multiple views of the tree geometries with their scores, and a large dataset of images with scores suitable for training deep neural networks. 
    more » « less
  2. Abstract

    In high seismic risk regions, it is important for city managers and decision makers to create programs to mitigate the risk for buildings. For large cities and regions, a mitigation program relies on accurate information of building stocks, that is, a database of all buildings in the area and their potential structural defects, making them vulnerable to strong ground shaking. Structural defects and vulnerabilities could manifest via the building's appearance. One such example is the soft‐story building—its vertical irregularity is often observable from the facade. This structural type can lead to severe damage or even collapse during moderate or severe earthquakes. Therefore, it is critical to screen large building stock to find these buildings and retrofit them. However, it is usually time‐consuming to screen soft‐story structures by conventional methods. To tackle this issue, we used full image classification to screen them out from street view images in our previous study. However, full image classification has difficulties locating buildings in an image, which leads to unreliable predictions. In this paper, we developed an automated pipeline in which we segment street view images to identify soft‐story buildings. However, annotated data for this purpose is scarce. To tackle this issue, we compiled a dataset of street view images and present a strategy for annotating these images in a semi‐automatic way. The annotated dataset is then used to train an instance segmentation model that can be used to detect all soft‐story buildings from unseen images.

     
    more » « less
  3. The PoseASL dataset consists of color and depth videos collected from ASL signers at the Linguistic and Assistive Technologies Laboratory under the direction of Matt Huenerfauth, as part of a collaborative research project with researchers at the Rochester Institute of Technology, Boston University, and the University of Pennsylvania. Access: After becoming an authorized user of Databrary, please contact Matt Huenerfauth if you have difficulty accessing this volume. We have collected a new dataset consisting of color and depth videos of fluent American Sign Language signers performing sequences ASL signs and sentences. Given interest among sign-recognition and other computer-vision researchers in red-green-blue-depth (RBGD) video, we release this dataset for use by the research community. In addition to the video files, we share depth data files from a Kinect v2 sensor, as well as additional motion-tracking files produced through post-processing of this data. Organization of the Dataset: The dataset is organized into sub-folders, with codenames such as "P01" or "P16" etc. These codenames refer to specific human signers who were recorded in this dataset. Please note that there was no participant P11 nor P14; those numbers were accidentally skipped during the process of making appointments to collect video stimuli. Task: During the recording session, the participant was met by a member of our research team who was a native ASL signer. No other individuals were present during the data collection session. After signing the informed consent and video release document, participants responded to a demographic questionnaire. Next, the data-collection session consisted of English word stimuli and cartoon videos. The recording session began with showing participants stimuli consisting of slides that displayed English word and photos of items, and participants were asked to produce the sign for each (PDF included in materials subfolder). Next, participants viewed three videos of short animated cartoons, which they were asked to recount in ASL: - Canary Row, Warner Brothers Merrie Melodies 1950 (the 7-minute video divided into seven parts) - Mr. Koumal Flies Like a Bird, Studio Animovaneho Filmu 1969 - Mr. Koumal Battles his Conscience, Studio Animovaneho Filmu 1971 The word list and cartoons were selected as they are identical to the stimuli used in the collection of the Nicaraguan Sign Language video corpora - see: Senghas, A. (1995). Children’s Contribution to the Birth of Nicaraguan Sign Language. Doctoral dissertation, Department of Brain and Cognitive Sciences, MIT. Demographics: All 14 of our participants were fluent ASL signers. As screening, we asked our participants: Did you use ASL at home growing up, or did you attend a school as a very young child where you used ASL? All the participants responded affirmatively to this question. A total of 14 DHH participants were recruited on the Rochester Institute of Technology campus. Participants included 7 men and 7 women, aged 21 to 35 (median = 23.5). All of our participants reported that they began using ASL when they were 5 years old or younger, with 8 reporting ASL use since birth, and 3 others reporting ASL use since age 18 months. Filetypes: *.avi, *_dep.bin: The PoseASL dataset has been captured by using a Kinect 2.0 RGBD camera. The output of this camera system includes multiple channels which include RGB, depth, skeleton joints (25 joints for every video frame), and HD face (1,347 points). The video resolution produced in 1920 x 1080 pixels for the RGB channel and 512 x 424 pixels for the depth channels respectively. Due to limitations in the acceptable filetypes for sharing on Databrary, it was not permitted to share binary *_dep.bin files directly produced by the Kinect v2 camera system on the Databrary platform. If your research requires the original binary *_dep.bin files, then please contact Matt Huenerfauth. *_face.txt, *_HDface.txt, *_skl.txt: To make it easier for future researchers to make use of this dataset, we have also performed some post-processing of the Kinect data. To extract the skeleton coordinates of the RGB videos, we used the Openpose system, which is capable of detecting body, hand, facial, and foot keypoints of multiple people on single images in real time. The output of Openpose includes estimation of 70 keypoints for the face including eyes, eyebrows, nose, mouth and face contour. The software also estimates 21 keypoints for each of the hands (Simon et al, 2017), including 3 keypoints for each finger, as shown in Figure 2. Additionally, there are 25 keypoints estimated for the body pose (and feet) (Cao et al, 2017; Wei et al, 2016). Reporting Bugs or Errors: Please contact Matt Huenerfauth to report any bugs or errors that you identify in the corpus. We appreciate your help in improving the quality of the corpus over time by identifying any errors. Acknowledgement: This material is based upon work supported by the National Science Foundation under award 1749376: "Collaborative Research: Multimethod Investigation of Articulatory and Perceptual Constraints on Natural Language Evolution." 
    more » « less
  4. Extracting roads in aerial images has numerous applications in artificial intelligence and multimedia computing, including traffic pattern analysis and parking space planning. Learning deep neural networks, though very successful, demands vast amounts of high-quality annotations, of which acquisition is time-consuming and expensive. In this work, we propose a semi-supervised approach for image-based road extraction where only a small set of labeled images are available for training to address this challenge. We design a pixel-wise contrastive loss to self-supervise the network training to utilize the large corpus of unlabeled images. The key idea is to identify pairs of overlapping image regions (positive) or non-overlapping image regions (negative) and encourage the network to make similar outputs for positive pairs or dissimilar outputs for negative pairs. We also develop a negative sampling strategy to filter false negative samples during the process. An iterative procedure is introduced to apply the network over raw images to generate pseudo-labels, filter and select high-quality labels with the proposed contrastive loss, and re-train the network with the enlarged training dataset. We repeat these iterative steps until convergence. We validate the effectiveness of the proposed methods by performing extensive experiments on the public SpaceNet3 and DeepGlobe Road datasets. Results show that our proposed method achieves state-of-the-art results on public image segmentation benchmarks and significantly outperforms other semi-supervised methods.

     
    more » « less
  5. Recent years have witnessed the great success of deep learning models in semantic segmentation. Nevertheless, these models may not generalize well to unseen image domains due to the phenomenon of domain shift. Since pixel-level annotations are laborious to collect, developing algorithms which can adapt labeled data from source domain to target domain is of great significance. To this end, we propose self-ensembling attention networks to reduce the domain gap between different datasets. To the best of our knowledge, the proposed method is the first attempt to introduce selfensembling model to domain adaptation for semantic segmentation, which provides a different view on how to learn domain-invariant features. Besides, since different regions in the image usually correspond to different levels of domain gap, we introduce the attention mechanism into the proposed framework to generate attention-aware features, which are further utilized to guide the calculation of consistency loss in the target domain. Experiments on two benchmark datasets demonstrate that the proposed framework can yield competitive performance compared with the state of the art methods. 
    more » « less