Joint Geometry & Scene Parsing

Depth (Scene geometry) and semantics are heavily correlated, this project (CVPR 2015) is trying to infer difference information in a unified framework in order to make different information benefit each other.  

Ours end to end training for semantic is different with that of Fully Conv Net. We show that we can only apply last fc7 to prediction each pixel of the image, which means the output resolution is not limited by pooling.

The depth prediction results from our model over NYU v2 data (654 images) can be downloaded here

Evaluation in an Understandable Way

In long term  we upgrade our algorithm but care not much about evaluations, in this project (WACV 2015),  we consider break out some factors that influence segmentation performance. From these factors we can understand more on the algorithms.  We have our code and model available here

Cascade Aethetic Image Cropping

In this project (WACV 2015), we developped effiecient cropping for better aethetic image.  (20x faster than previoius state-of-the-art and a much better performance) 

We apply cascade model and cascade efficient sliding window for inference. 

Visual Recognition for Image Classification

A project aiming at learning compact low level image descriptor representation. 

Some progress obtained (CVPR 2013), while still much is unknown: 

How to select optimum parameters such as patch size, handle 3D variation, extent it efficiently to hierarchic model.....

Salient Object Detection in Web Images (Data)

Aim at helping web image search to do image cropping or deeper description of images.

It determine whether an image contains a salient object and localizing the salient object  or salient regions with our constructed global saliency descriptor (CVPR2012). 



How to learn semantic salient regions for human without losing generalization is still an opening question. 

Superpixel Image Scene Parsing

Aim at parsing images with structure information (saliency guided or geometric guided) and finally do semantic segmentation.

Results in a structure consistent superpixel segmentation (ICCV 2011, IJCV 2013 )

  

Image Dominant Color Extraction

Aim at extracting the human perceptional dominant color for bing's color filter image search of Microsoft (MM' 12 Demo). Facilitate searching the objects of desired color. 

Localize the salient object,  mapping each pixel inside softly to the proposed color names and extract regions considering context information. (ICMEW 2012).