r-cnn spp-net bing

[bhack]

I reopen this from #482 because r-cnn was merged and ticked closed. But there was some interest in trying BING (now is available) because r-cnn still need a closed source matlab candidate region generator.
SPP-NET are defined here: http://arxiv.org/abs/1406.4729v1

[kloudkl]

Thanks for sharing the link! #482 is in Python and only serves for tutorial purpose. The R-CNN paper follows the four steps in the standard object detection pipeline, i.e. region of interest selection, feature extraction of each region, object classification of each region, and region fusion. For practical applications, these steps must all be implemented in C++.

This issue is more likely to be an umbrella of a series of issues with smaller scopes.

1. General object ROI selection API and some concrete implementations or adapters of external implementations such as BING.
2. SPP-net which is a highly efficient feature extractions method for object detection and seems very similar to DenseNet feature pyramid computation #308 DenseNet feature pyramid computation.
3. A first class module for an end-to-end object detection pipeline with detailed documents, off-the-shelf trainer and detector examples, and pre-trained models for common objects, e.g. faces etc.

[shelhamer]

This thread has been dead for some time now, but if an R-CNN is trained on an alternative proposal method with the same or better performance then a PR to update the detection example would be welcome.