Accuracy and Usability of Smartphone-Based Distance Estimation Approaches for Visual Assistive Technology Development.

Goal: Distance information is highly requested in assistive smartphone Apps by people who are blind or low vision (PBLV). However, current techniques have not been evaluated systematically for accuracy and usability. Methods: We tested five smartphone-based distance-estimation approaches in the image center and periphery at 1-3 meters, including machine learning (CoreML), infrared grid distortion (IR_self), light detection and ranging (LiDAR_back), and augmented reality room-tracking on the front (ARKit_self) and back-facing cameras (ARKit_back). Results: For accuracy in the image center, all approaches had <±2.5 cm average error, except CoreML which had ±5.2-6.2 cm average error at 2-3 meters. In the periphery, all approaches were more inaccurate, with CoreML and IR_self having the highest average errors at ±41 cm and ±32 cm respectively. For usability, CoreML fared favorably with the lowest central processing unit usage, second lowest battery usage, highest field-of-view, and no specialized sensor requirements. Conclusions: We provide key information that helps design reliable smartphone-based visual assistive technologies to enhance the functionality of PBLV.

Training AI to Recognize Objects of Interest to the Blind and Low Vision Community.

Recent object detection models show promising advances in their architecture and performance, expanding potential applications for the benefit of persons with blindness or low vision (pBLV). However, object detection models are usually trained on generic data rather than datasets that focus on the needs of pBLV. Hence, for applications that locate objects of interest to pBLV, object detection models need to be trained specifically for this purpose. Informed by prior interviews, questionnaires, and Microsoft's ORBIT research, we identified thirty-five objects pertinent to pBLV. We employed this user-centric feedback to gather images of these objects from the Google Open Images V6 dataset. We subsequently trained a YOLOv5x model with this dataset to recognize these objects of interest. We demonstrate that the model can identify objects that previous generic models could not, such as those related to tasks of daily functioning - e.g., coffee mug, knife, fork, and glass. Crucially, we show that careful pruning of a dataset with severe class imbalances leads to a rapid, noticeable improvement in the overall performance of the model by two-fold, as measured using the mean average precision at the intersection over union thresholds from 0.5 to 0.95 (mAP50-95). Specifically, mAP50-95 improved from 0.14 to 0.36 on the seven least prevalent classes in the training dataset. Overall, we show that careful curation of training data can improve training speed and object detection outcomes. We show clear directions on effectively customizing training data to create models that focus on the desires and needs of pBLV.Clinical Relevance- This work demonstrated the benefits of developing assistive AI technology customized to individual users or the wider BLV community.

NYU-VPR: Long-Term Visual Place Recognition Benchmark with View Direction and Data Anonymization Influences.

Visual place recognition (VPR) is critical in not only localization and mapping for autonomous driving vehicles, but also assistive navigation for the visually impaired population. To enable a long-term VPR system on a large scale, several challenges need to be addressed. First, different applications could require different image view directions, such as front views for self-driving cars while side views for the low vision people. Second, VPR in metropolitan scenes can often cause privacy concerns due to the imaging of pedestrian and vehicle identity information, calling for the need for data anonymization before VPR queries and database construction. Both factors could lead to VPR performance variations that are not well understood yet. To study their influences, we present the NYU-VPR dataset that contains more than 200,000 images over a 2km×2km area near the New York University campus, taken within the whole year of 2016. We present benchmark results on several popular VPR algorithms showing that side views are significantly more challenging for current VPR methods while the influence of data anonymization is almost negligible, together with our hypothetical explanations and in-depth analysis.