As the price of photography and video equipment drops while availability improves, visual data from the public is becoming the most abundant source of wildlife data. ÌýHowever, curating this high volume of data from "citizen science" and social media poses new scalability challenges for both researchers and computer scientists alike. ÌýThis talk introduces the Wildbook platform -- an open-source web-based project (wildbook.org) -- that leverages a suite of deep learning tools to automatically process high volumes of image data for conservation. ÌýThe platform uses Convolutional Neural Networks (CNNs) on several computer vision tasks, including: image classification, bounding box regression, instance classification, class segmentation, and object of interest classification. ÌýOur deep learning stack utilizes Theano and PyTorch on the NVIDIA's CUDA, CNMeM, and CuDNN deep learning stack and employs multiple Titan V GPUs for efficient processing. ÌýFurther, instance recognition allows Wildbook to individually identify individual animals through time using a pipeline of SIFT features to find distinctive patches, Approximate Nearest Neighbors to find visual neighbors, and LNBNN to associate and re-rank visual match correspondences. ÌýA random forest pair-wise match verifier is then used to curate animal sightings for human-in-the-loop population curation. ÌýLastly, our computer vision pipeline works alongside an intelligent agent that can automatically ingest video data from YouTube using NLP and OCR with Azure Cognitive Services. Ìý We present our conservation work for wildlife across the globe in the context of the latest advances in deep learning and how Wildbook is helping to convert wildlife conservation into a data-driven science.Ìý
Jason Parham received his B.S. in Computer Science / Mathematics from Pepperdine University in Malibu, CA in 2008 and holds a M.S. in Computer Science from RPI in Troy, NY. ÌýJason is finishing his Ph.D. as a candidate under the advising ofÌýDr. Charles Stewart at RPI. ÌýJason’s Masters thesis was on the design and implementation ofÌýa citizen science-powered photographic censusing of the zebra and giraffe in the Nairobi National Park. ÌýHis current doctoral research focuses on animal detection and classification, using deep learning on wildlife imagery, to power automated photographic censusing. ÌýJason is a co-developer ofÌýWildbook‘s Image Analysis components, which are used to monitor animal populations in conservancies around Kenya and which integrate with the Wildbook data management platform for a suite of species. Previously, Jason worked three years for Kitware, Inc. in Clifton Park on detecting vehicles and military aircraft in overhead satellite imagery. ÌýBelow are some select publications: - An Animal Detection Pipeline for Identification Lake Tahoe, CA. IEEE WINTER CONFERENCE ON APPLICATIONS OF COMPUTER VISION (WACV 2018) Mar 2018
- Wildbook: Crowdsourcing, Computer Vision, and Data Science for Conservation New York, NY. BLOOMBERG DATA FOR GOOD EXCHANGE 2017 Sep 2017
- Animal Population Estimation Using Flickr Images Troy, NY ACM WEB SCIENCE CONFERENCE 2017 (WEBSCI‘17). INTERNATIONAL WORKSHOP ON THE SOCIAL WEB FOR ENVIRONMENTAL AND ECOLOGICAL MONITORING (SWEEM 2017)
- Efficient Generation of Image Chips for Training Deep Learning Networks Anaheim, CA. PROCEEDINGS OF SOCIETY OF PHOTO-OPTICAL INSTRUMENTATION ENGINEERS (SPIE) · AUTOMATIC TARGET RECOGNITION (ATR). XXVII Apr 2017.
Ìý