We have launched!!! Since 2013, I have been
the lead scientist at Amazon Go,
using computer vision and machine learning to re-invent retail.
We have just launched our first "Just-Walk-Out" store that automatically figures out purchases without customer effort, completely eliminating check-out (that "unnecessary" and annoying wait).
I was a research scientist at Intel Labs during 2008-2013, working closely with faculty and students at University of Washington. Prior to sunny Seattle, I was a research
assistant professor at the Toyota Technological Institute at Chicago (TTI-C). I received my Ph.D.
from U.C. Berkeley in 2006, under the supervision of Jitendra Malik.
Code available: I have put together C++ implementations of both our
kernel descriptor features (NIPS10) and the more recent hierarchical sparse coding
features (NIPS11) in a live webcame demo, here released under BSD license. What's more, the demo is now running on an Android phone!!
Progress on contour detection: move beyond Pb and use sparse codes to compute local oriented gradients. F=0.74 (up from 0.71 of gPb) on BSDS500, a large step forward (human=0.80). Great RGB-D results: F=0.62 (vs gPb 0.53) on NYU Depth (v2).
Check out the demo video for our Ubicomp paper on fine-grained kitchen activity recognition and tracking. News article at New Scientist.
A C++ implementation of our kernel descriptor features (NIPS10, IROS11),
~5 times faster than the matlab version, and with a live demo using webcams. Please try
this out and let me know your comments and suggestions.
Upcoming papers at ISER, Ubicomp and the Robotics and Automation Magazine (RAM).
CVPR paper accepted on scene labeling on both RGB-D (indoor) and image-only (outdoor) scenes. Preprint available.
Liefeng's NIPS paper on hierarchical orthogonal matching pursuit for learning image features.
The (Matlab) code for kernel descriptors is now available. Please try it out!!
Check out our ubicomp final video on interactive mapping on youtube.
Co-organized the 2nd
RGB-D workshop at RSS on advanced perception using depth cameras; it's a great success!! 18 presentations, 7 demos and over 70 attendees. All the papers, videos, slides will be available online.
Two BMVC papers to appear on material recognition and video segmentation.
Two IROS papers to appear on depth kernel descriptors and object discovery based on scene changes.
Our interactive mapping work (online user interaction in real-time mapping) to appear at Ubicomp 2011.
Histograms of Sparse Codes for Object Detection.
Xiaofeng Ren and
Deva Ramanan, at CVPR 2013.
Move beyond HOG! Use learned sparse code dictionaries to significantly improve object detection accuracy
Multipath Sparse Coding Using Hierarchical Matching Pursuit.
[abstract][pdf][code] Liefeng Bo, Xiaofeng Ren and
Dieter Fox, at CVPR 2013.
Extend hierarchical matching pursuit (NIPS11) to a reconfigurable architecture that captures structures of varying scale and deformation
RGB-D Object Discovery via Multi-Scene Analysis.
[pdf] Evan Herbst, Xiaofeng Ren and Dieter Fox, at IROS 2011.
Enable a robot to automatically discover and cluster objects via multiple visits to a scene; works on all objects
Sparse Distance Learning for Object Recognition Combining RGB and Depth Information.
(best vision paper)
[abstract][pdf] Kevin Lai, Liefeng Bo, Xiaofeng Ren and Dieter Fox, at ICRA 2011.
Local distance learning and feature selection using instance-to-class distance
Toward Object Discovery and Modeling via 3-D Scene Comparison.
[abstract][pdf] Evan Herbst, Xiaofeng Ren and Dieter Fox, at ICRA 2011.
How can a robot discover objects robustly? By visiting a scene and finding out the changes
Discriminative Mixture-of-Templates for Viewpoint Classification.
[abstract][pdf] Chunhui Gu and Xiaofeng Ren, at ECCV 2010, Crete, Greece, 2010.
First paper on discriminative models for viewpoint/pose recognition, large improvement: 57%=>74% on 3DObject
Manipulator and Object Tracking for In Hand Model Acquisition.
[pdf] Michael Krainin, Peter Henry, Xiaofeng Ren and Dieter Fox, at the Mobile Manipulation and Best Practices in Robotics Workshops at ICRA 2010.
Figure-Ground Segmentation Improves Handled Object Recognition in Egocentric Video.
Xiaofeng Ren and Chunhui Gu, at CVPR 2010, San Francisco, 2010.
Egocentric recognition can work!! ~90% accuracy on a very challenging dataset for objects-in-hand. Check out the videos.
Multi-Scale Improves Boundary Detection in Natural Images.
Xiaofeng Ren, in ECCV '08, Marseille, 2008.
Multi-scale does help contour detection on natural images - extensive benchmarking and analysis