Any faster feature extraction tools (better than the one in bottom-up-detection and lxmert)?

Hi @ChenRocks @linjieli222,
It is still ok even if it is not used by [VisualBERT](https://github.com/uclanlp/visualbert/), [LXMERT](https://github.com/airsplay/lxmert), and [UNITER](https://github.com/ChenRocks/UNITER).

- I have to handle 1,102,076+ personal images and have to find a fast one even the performance might be lower -- it is fine NOT to be any of [VisualBERT](https://github.com/uclanlp/visualbert/), [LXMERT](https://github.com/airsplay/lxmert), and [UNITER](https://github.com/ChenRocks/UNITER), but **better faster to implement**.

For example, by comparison with one GPU (cpu-only is presumably [not tolerable](https://github.com/airsplay/lxmert/issues/89)), for NLVR2 107,292 images, [lxmert](https://github.com/airsplay/lxmert) takes 5-6 hours to extract faster-rcnn features by [this caffe](https://github.com/peteanderson80/bottom-up-attention/tree/master/caffe).
I also follow visualBERT's [issue#1](https://github.com/uclanlp/visualbert/issues/10) and [issue#10](https://github.com/uclanlp/visualbert/issues/10), [LXMERT](https://github.com/airsplay/lxmert), [transformers-VQA](https://github.com/YIKUAN8/Transformers-VQA/issues/5).

No repository file or test is named. Start by reviewing the linked bottom-up-attention, VisualBERT, LXMERT, and UNITER feature-extraction approaches, then compare their one-GPU runtime for the stated image volume. Done would require a selected or implemented faster extraction approach with benchmark results and documented trade-offs.

Written by the indexing model from the issue text.