Learning 3-D Scene Structure from a Single Still Image.

Home | Monocular Depth Estimation | Stanford Range Image Data | Single Image 3-D Reconstruction

Full documentation of results and human evaluation

We tested our mode on 134 images collected using our 3- d scanner setup, and also on 588 images downloaded from internet. The image on the internet were collected by issuing keywords on Google image search.

To collect data and to perform the evaluation of the algorithms in a completely unbiased manner, a person not associated with the project was asked to collect images of environments (greater than 800x600 size). The person chose the following keywords to collect the images: campus, garden, park, house, building, college, university, church, castle, court, square, lake, temple, scene.


The following links are the results of our algorithm and HEH's in vrml format, categorize by keywords issed in google image search.

building campus castle church college
court garden house park lake
scene square temple university
Typical Result download

Our algorithm's result is given by Keyword_index.wrl and HEH photo Popup result will be index by Keyword_index_HEH.wrl for comparison. Note: These images were collected using Google images. If one of the images in the data belongs to you; then please email anonymous.

Full documentation of the evaluation of the algorithms, is available in the Excel files.


You will need a VRML viewer. We suggest Cortona plugin for Internet explorer in Windows. Browser plugin takes less than a minute to install.
For Linux, you can try this.

Training set of 400 images and laser range data

We used 400 images (size of 1704x2272) and laser range data (size of 55x305 measurement points) as training data to learn depths from image features.