Csordás Róbert

As far as I know, TF has only a CPU based NMS implementation. I don't think so that it can be parallelized so well that it can get any significant...

Dear philokey, This is a non hyper-parameter optimised version of the code. Unfortunately I cannot publish the optimised version, because my employer would be not happy to share this information...

I checked the most similar version after hyperparameter tuning, and it is ~84% on Kitti Car benchmark (it is trained on Kitti, too), which is very good for an RFCN.

Supporting batch of >1 is not trivial because of varying number of bounding boxes per image, and varying image size. Boxes could be represented by TensorArrays, or a padded tensor...

Training on your own data that is different from MS COCO is not trivial. You need to write your own data loader. You can use the Dataset/CocoDataset.py as an example/reference....

It was trained for 2.2M iterations. After 100-200k iters enable training of the base network (trainFrom argument, you can experiment with different values). After 1.5M iterations decrease the learning rate...

Thank you qiningonline! I was very busy recently and not able to support the issue.

The problem is that training with batch size of >1 is not supported. Therefore it can't be parallelized efficiently. Supporting batch of >1 is not trivial because of varying number...

Could you attach a file where this happens? Just about the filename: are those notes, or a pdf annotated by the reader app? Because the tool works for the reader...

The file has wrong format, it doesn't terminate some streams it uses. However, this doesn't seem to bother the PDF viewer programs. The PDF parser library is not mine, so...