Episode Summary
Executive Summary: Qualcomm’s Fatih Purikli discusses his path between academia and industry and three CVPR papers advancing perception: a transformer-based panoptic segmentation system, self-supervised optical flow with occlusion reasoning, and dense vision transformers for inverse rendering. He emphasizes flexible, end-to-end attention mechanisms, robustness, and efficient deployment for real-world applications like autonomous driving, XR, robotics, and mobile devices.
Main Topics: Fatih Purikli’s background and research focus (Priority: 5/5): Purikli describes moving between academia and industry multiple times, with a long-standing focus on computer vision and perception as the basis for how humans and machines understand the world. Panoptic segmentation with relational context encoding (Priority: 5/5): The first paper combines instance and semantic segmentation into one end-to-end transformer-based system that learns relations across pixels and instances, improving accuracy across many segmentation backbones. Self-supervised optical flow with occlusion handling (Priority: 5/5): The second paper reduces dependence on large labeled datasets by using synthetic transformations and occlusion masks to train optical flow models more robustly and accurately. Dense vision transformers for inverse rendering (Priority: 4/5): The third paper uses attention to infer hidden scene properties from a single indoor image, such as lighting, reflectance, room layout, and materials, enabling realistic scene editing and relighting. Wireless AI perception and omnidirectional vision workshops (Priority: 4/5): Qualcomm is also advancing perception through CVPR workshops on wireless sensing and omnidirectional cameras, highlighting multimodal perception and 360-degree visual understanding. Deployment and efficiency for real-world use (Priority: 4/5): Across all projects, Purikli stresses making models efficient enough for mobile, real-time, and production settings, especially on Qualcomm platforms.
Key Arguments: Computer vision is central because visual perception dominates human understanding of the world, and AI can now learn these capabilities from data instead of hand-crafted mathematical rules. Panoptic segmentation is hard because it must jointly identify countable instances and uncountable semantic regions while separating overlapping objects. Transformer-style attention improves segmentation by focusing on the most relevant parts of the image and can be integrated end-to-end so instance and semantic branches support each other. The panoptic method is backbone-agnostic and improved results across more than 20 segmentation algorithms, showing the generality of the approach. Optical flow training is limited by the lack of real labeled motion ground truth, so self-supervision via synthetic transformations and occlusion reasoning is a practical alternative. Inverse rendering benefits from attention because it must infer invisible scene properties from visible cues, and shared reasoning across tasks improves realism and editability. Efficiency matters as much as accuracy: Qualcomm is working to make these perception models run in real time on mobile and extend them to multi-camera and higher-dimensional settings. The field is not solved; better robustness to noise, blur, occlusion, domain shift, and new object classes is still needed before these systems are broadly product-ready.
Data Points: Computer activity dedicated to visual perception: 70-75% - Purikli cites this as a reason vision is so central to intelligence and perception research. Number of segmentation algorithms tested: More than 20 - The panoptic segmentation paper reportedly improved results when plugged into more than 20 segmentation pipelines. Benchmark scale for optical flow: More than 200 solutions - Purikli says the KITTI and Sintel benchmarks had over 200 solutions competing. Deep learning solutions on optical flow benchmarks: 100+ - He notes that more than 100 of the benchmark solutions were AI/deep learning-based. Localization precision for Wi-Fi sensing: Less than 10 cm - He mentions Wi-Fi-based sensing can track person movement with sub-10-centimeter accuracy depending on access points. CVPR workshop count for wireless AI perception: First time - He frames the wireless AI perception workshop as the first workshop of its kind at CVPR. Omnidirectional vision workshop edition: Third time - He says the omnidirectional computer vision workshop is in its third edition. Demo count from Qualcomm perception team: 2 - He says his team is bringing two demos: OXIDAP and 4K image super-resolution.
Pivotal Quotes: "Vision is the way that we understand, make sense of the world, life, everything around us." — Fatih Purikli: Explaining why computer vision has always been his core research interest. "The same network can do in an end-to-end fashion these two tasks together." — Fatih Purikli: Describing the panoptic segmentation approach that jointly handles instance and semantic segmentation. "We do not need such a big data set. We will do self-supervised learning, unsupervised learning." — Fatih Purikli: Summarizing the optical flow paper’s strategy for reducing dependence on large labeled datasets.
Implications: The conversation signals that perception is moving toward unified, attention-driven systems that are more data-efficient, robust, and deployable. These advances could improve autonomy, XR, robotics, and mobile sensing while opening new multimodal and 360-degree perception applications.