search

forresti / caffe

Caffe
http://caffe.berkeleyvision.org/
Other
5 stars 2 forks source link

ask for quantitative results on voc pascal dataset #9

Open zhangliliang opened 9 years ago

zhangliliang commented 9 years ago

Hi, does someone has run DenseNet on VOC detection dataset and get some quantitative results

moskewcz commented 9 years ago

if you mean specifically with this codebase, i think the answer is no. however, i think there are now various recent published papers that use DCNNs (some using caffe) for object detection on the PASCAL VOC dataset(s), and i think some/many of them are doing similar things to densenet in terms of using 'dense' CNN features -- as one might expect, since it's a pretty straighforward/simple thing to do, albeit with various caveats ...

the most closely related/derived code/work to this one that i'm aware of is this one from ross/forrest; see ross's homepage: http://www.cs.berkeley.edu/~rbg/ paper: http://arxiv.org/pdf/1409.5403v2.pdf code: https://github.com/rbgirshick/DeepPyramid

of course, you should also look at the 'regular' rcnn code and papers as well, as they're certainly quite related.

also, FWIW, i'm still hoping to get some form of simple detection example(s) (based on this densenet code and/or on some other existing or derived codebases) upstreamed into caffe at some point. it's still an open question what functionaliy belongs inside caffe versus what are things that can be layered on top. further, there are various technical details and python/matlab interface issues to deal with.