Computer Vision Technical Talks at Motionlab Berlin
Details
"Leaving No Pixels Behind: Deep Learning for Perfect Cutouts"
Speaker: Imran Kocabiyik, withoutbg
Removing backgrounds from images is a challenging task, even for advanced deep learning models. The human eye is highly sensitive to minor imperfections, making high-quality outcomes crucial. In this talk, Imran Kocabiyik will demonstrate how withoutbg achieves clean, natural-looking image extractions while addressing the issues of costly training data and the need to handle diverse image types. Their approach effectively balances intelligent model design and meticulous data selection, resulting in impressive performance suited for real-world applications.
"AI on the Dance Floor: Multimodal Segmentation of Choreography Videos"
Speaker: Dr. Paras Mehta, sylby
Ever struggled to learn a dance routine by constantly rewinding YouTube videos? In this talk, Paras presents an approach based on temporal convolutional networks and pose estimation to automatically segment choreography videos into individual moves by leveraging both audio and visual modalities.
"EnvisionHGdetector: A Framework for Detecting and Analyzing Hand Gestures During Speech"
Speaker: Sharjeel Shaikh, University of Potsdam, HPI
We present EnvisionHGdetector, a toolkit for studying hand movements during speech. It measures hand motion, compares gestures, and labels gesture segments using Mediapipe tracking and a custom neural network. Tested on over 8,000 gestures, it achieved approximately 75% accuracy. We also discuss plans to improve accessibility for gesture researchers.
"When Images Look Alike: Intro to Dataset Curation"
Speaker: Antonio Rueda-Toicen
This talk introduces dataset curation in computer vision, focusing on visually similar images. We discuss use cases in vacation rental search and art recommendations. We demonstrate how Voxel51 helps identify image similarity, improving data quality and model reliability.
Registration
Please register through Voxel51's page to confirm your attendance.