mohhao

I can't find the paper "LightNet: Light-weight Networks for Semantic Image Segmentation",either,really appreciate the work you did, can you offer a paper site or do you have a plan for...

denseCRFs needn't training

> Hi, > I am using the trained model (as provided in the medium blog), and when I use it on test data, I do not get any correct predictions....

> I had the same problem, did anyone solve it or know how to solve it ?Thanks a lot > > Best regard Hi, have you solved it?

> > > I had the same problem, did anyone solve it or know how to solve it ?Thanks a lot > > > Best regard > > > >...

> > > > > I had the same problem, did anyone solve it or know how to solve it ?Thanks a lot > > > > > Best regard...

> In [detector.py](https://github.com/TropComplique/light-head-rcnn/blob/master/detector/detector.py), your possitive anchors are defined as `is_positive_anchor = tf.greater_equal(anchor_matches, 0)`[Line 138](https://github.com/TropComplique/light-head-rcnn/blob/cac2fab53964e8a9057351d7c85c2f6537c2e9cc/detector/detector.py#L138), so don't include negetive sample. You use it in function **positive_negative_subsample**, in this [function](https://github.com/TropComplique/light-head-rcnn/blob/cac2fab53964e8a9057351d7c85c2f6537c2e9cc/detector/losses_and_subsampling.py#L53) mistake appears...