Clockwork FCNs for Fast Video Processing

ABOUT THE PROJECT

At a glance

Video-rate semantic segmentation with contemporary convolutional networks requires considerable computational resources and suffers degraded performance from the image-blur that results from rapidly moving cameras or objects. This project will investigate complementary innovations to effect fast and accurate labeling of blur-free videos of dynamic scenes. While pixels may change rapidly from frame to frame, the high-level semantic content of a scene is more stationary. The team will use a novel, frame-asynchronous, “clockwork” convnet, which pipelines processing over time with different network update rates at different levels of the semantic processing hierarchy, significantly reducing frame processing time with little or no degradation in performance. The team’s proposed approach will use learned spatial semantic transitions to reduce redundant processing to improve the overall recognition performance. Further, the team will address the problem of motion blur by leveraging an overlooked feature of laser-based projector-camera systems: individual pixels of a scene are individually laser-illuminated so briefly that blur is nearly non-existent. It has been observed that an unmodified low-power pico-projector suffices as an indoor illumination source to yield blur free video with a consumer-grade CMOS camera. The clockwork FCN provides faster and accurate semantic segmentation of video while the combination of network and projector-camera hardware opens up the possibility of processing videos with extreme motion as captured on assembly lines or by flying drone.

The team’s new approach is the first reported design for blur-free imaging despite extreme motion through the use of a laser-projector-camera system.
 
principal investigatorsresearchersthemes
Trevor DarrellJ. Hoffman
Kate Rakelly
E. Shelhamer
Semantic Segmentation
Convolutional Networks

 

BAIR/CPAR/BDD Internal Weekly Seminar

Event Location: 
250 Sutardja Dai Hall

The Berkeley Artificial Intelligence Research Lab co-hosts a weekly internal seminar series with the CITRIS People and Robots Initiative and the Berkeley Deep Drive Consortium. The seminars are every Friday afternoon in room 250 Sutardja Dai Hall from 3:10-4:10 PM, and are open to BAIR/BDD faculty, students, and sponsors. Seminars will be webcast live and recorded talks will be available online following the seminar. 

Schedule:

http://citris-uc.org/bair-seminar-series/