Bridging the Gap between Computer Vision Research and Real-World Data Processing

<p dir="ltr">This thesis focuses on bridging the gap between computer vision research and real-world data and applications. We will present our findings and contributions through two application-oriented projects — quality estimation for recognition tasks and in-the-wild camera trap image processing.In the first part, we recognize that image quality is an important factor that influences the outcome of recognition tasks, such as face recognition or person re-identification with full-body images. Using low quality inputs could yield inaccurate results which may not be useful for further tasks. Therefore, we propose an unsupervised quality metric for person re-identification task by measuring consistency under perturbations. To gain more insight to quality metrics for recognition tasks, we also conducted experiments to examine face quality metrics’ weaknesses under stress tests.The second part discusses the significant gap between curated dataset and real-world application it represents using the example of camera trap image processing.In this project, we need to design a system that detects and classifies animal species from in-the-wild camera trap videos for ecologists. First, using the animal detection / localization task, we illustrate the dramatic difference between un-curated data and curated data in terms of false animal detections. While data curation removes some undesirable data, it often removes useful data as well. We also proposed methods and evaluation protocols for measuring data difference, as well as a specific tool to reduce false detections for camera trap images. Then, we tackle the classification of unseen species using few-shot learning. In doing so, we demonstrate that good performance on benchmark data does not translate to real-world data, and offer practitioners guidelines on choosing effective component to build a realistic system. Overall, our processing system is designed so that the practitioners can adjust the system and achieve a balance between accuracy and amount of data rejected. </p><p dir="ltr"><br></p>