Applied ML seminars – Anomaly Segmentation (אירוע)
אני שמח להזמין אתכם לאירוע השלישי שלנו בסדרת Applied ML seminars שנעשים בשיתוף עם Applied Materials. בכל אירוע מסוג זה, ניקח נושא אחד שמעניין את הקהילה ונדבר עליו בהרחבה מכמה זוויות שונות. באירוע הזה נדבר על מקרה שמאוד נפוץ בקרב חברי הקהילה: מה עושים אם יש לנו טעויות בתיוגים, דאטה מועט או אפילו סתם דאטה לא מאוזן בצורה משמעותית. אלו בעיות שכל אחד נתקל בהם במהלך העבודה שלו ובאירוע זה יהיו לנו שלושה דוברים שידברו על האתגרים האלו בבעיות סגמנטציה (כמובן שניתן ללמוד מזה על עולמות תוכן נוספים).
האירוע יתקיים באופן וירטואלי ב-1.3 בשעה 18:00 ויכלול שלושה דוברים מחברות שונות.
ההקלטה המלאה של האירוע:
#1 Title: Making a lemonade out of any label: Semantic Segmentation using Noisy Annotations
Lecturer: Sveta Paster, Algorithm developer at Applied Materials.
Abstract:
Applied Materials is a worldwide company that develops and manufactures equipment used in the wafer fabrication steps of creating a semiconductor device. The Israeli R&D center is mainly focusing on inspection, review and metrology solutions.
Wafer fabrication is complex and unstable process that suffers from variant types of defects which can harm chip functionality and yield. One of the divisions in AMAT Israel is EDC (electric beam defect control) which focuses on developing solutions for detection and classification of those defects in order to improve customers wafer yield.
A possible solution to defects detection is semantic anomaly segmentation based on deep-learning neural networks.
In order to train a fully-supervised semantic segmentation model there is a need for accurate labels, which requires a tedious and expensive human resource effort.
Moreover, in real life applications, annotations can be noisy, partial or absent.
One direction to address this problem is by extracting more information out of an image in order to improve labels.
This meetup introduces possible alternatives such as: refine annotations, creating pseudo labels from image level information, etc.
Labels, can’t live with them, can’t live without them!
#2 Title of the talk: Leveraging Pre-trained Vision Transformers for Part Co-Segmentation
Lecturer: Shir Amir, Graduate student in the Weizmann Institute of Science, advised by Dr. Tali Dekel.
Abstract: Vision Transformers (ViTs) have recently become ubiquitous in the Computer Vision domain. In contrast to CNNs, ViTs maintain feature's resolution throughout their layers and possess the entire input image as the receptive field. In this work, we investigate the deep features of a self-supervised pretrained transformer.
We demonstrate that such features encode powerful high-level information at high spatial resolution, i.e. capture semantic object parts; and the encoded semantic information is shared across related, yet different object categories, i.e. super-categories. These properties allow us to leverage ViT features for several applications including co-segmentation, part co-segmentation and correspondences – all achieved by applying only lightweight methodologies to deep ViT features.
Our part co-segmentation method achieves state-of-the-art results and can be readily applied to an arbitrary number of input images, ranging from as little as a pair of images to a collection containing thousands of images. This lightweight method can be especially useful in scenarios where the data is scarce and no human annotations are available. (see project page: https://dino-vit-features.github.io).
#3 Presentation Title: Finding Nothing with Segmentation
Lecturer: Eitan kassuto, Deep Learning researcher at Trax
Abstract:
Trax Retail is a leading operator in the RetailTech industry, which top FMCG manufacturers and retailers trust for quick and accurate insights at a large scale. Trax Retail utilizes computer vision algorithms as a core technology to provide an accurate, scalable solution for detecting and recognizing products of interest at a granular level. Actionable insights are generated in a reporting layer using the recognition results, usually in the form of KPI (key performance indicators) dashboards to help inform decision-makers.
Some of these KPIs rely on detecting Empty spaces on the shelf, a rare occurrence resulting in few samples. This is challenging when one must train, optimize and evaluate a model to distinguish Empty from other classes. This talk will discuss the process of how we designed and built a pipeline to train an accurate segmentation model despite this data imbalance and other constraints.
ניתן להירשם לאירוע דרך הלינק הזה.