ROS Real Time Inference

[akeaveny]

Hi, I've a few questions after looking at your ROS implementation!
Similar to you I'm using MaskRCNN w/ DenseFusion (but trained both on my own synthetic images).

I was able to implment my own ROS node after a day of configuring my conda env which required:

• a compatable version of MaskRCNN w/ Python2.7 &
• Merging my tensorflow 1.14.0 (MaskRCNN) and pytorch (DenseFusion) envs ..

I'm able to run this live w/ a ZED camera on a RGB-D image to get an inference time of ~2fps.
I run MaskRCNN and DenseFusion live on seperate GPUs & my setup is 2x GeForce RTX 2080 Ti's

I understand that your pipleline is more complex than grabbing a ZED RGB-D image-> Segmentation -> Pose Estimation but I'd curious as to how I can decrease inference time for MaskRCNN.

One thing I noticed with your repo is that you run MaskRCNN for e.g. 100 frames and save masks for DenseFusion.
I believe MaskRCNN is faster when operating on batch rather than one image per se.

Any feedback is appreciated!

[hoangcuongbk80]

Hi, It should be faster if you put multiple images into a batch and operate the batch. In addition, there is number of things to consider if you want to reduce time execution such as use a smaller backbone network, use smaller images, reduce the maximum number of instances per image, ...
Good luck!