Code for several state-of-the-art papers in object detection and semantic segmentation.
MatConvNet-Calvin is a wrapper around MatConvNet that (re-)implements several state of-the-art papers in object detection and semantic segmentation. This includes our own work "Region-based semantic segmentation with end-to-end training" [5]. Calvin is a Computer Vision research group at the University of Edinburgh (http://calvin.inf.ed.ac.uk/). Copyrights by Holger Caesar and Jasper Uijlings, 2015-2016.
git clone https://github.com/nightrome/matconvnet-calvin.git
cd matconvnet-calvin
git submodule update --init
cd matconvnet/matlab; vl_compilenn('EnableGpu', true); cd ../..;
cd matconvnet-calvin/matlab; vl_compilenn_calvin(); cd ../..;
setup();
downloadModel('frcn');
downloadModel('fcn');
downloadModel('e2s2_full');
downloadModel('e2s2_fast');
demo_frcn()
demo_fcn()
Method | Model | Class accuracy | Global accuracy | Mean IOU | Training epochs | Training time |
---|---|---|---|---|---|---|
This | FCN-16s | 48.8% | 83.8% | 36.7% | 50 | 12h |
[2] | FCN-16s | 51.7% | 85.2% | 39.5% | 175? | - |
demo_e2s2_full()
or demo_e2s2_fast()
Model | Class accuracy | Training epochs | Training time | GPU RAM |
---|---|---|---|---|
Full | 66.2% | 25 | 75h | 8.5GB |
Fast | 62.5% | 10 | 20h | 6.0GB |
[5] | 64.0% | 30 | - | - |
Except for [5], none of the methods implemented in MatConvNet-Calvin is authorized by the original authors. These are (possibly simplified) reimplementations of parts of the described methods and they might vary in terms of performance. This software is covered by the FreeBSD License. See LICENSE.MD for more details.
If you run into any problems with this code, please submit a bug report on the Github site of the project. For other inquiries contact holger-at-it-caesar.com.