How to let Tensorflow object detection api use gray image to train(just 1 channel for input tensor)?

Question

I just want to get real time speed when using model trained by tensorflow object detection api, the input tensor has shape[1, width, height,3], it is 3 channels,but I think if I can just use 1 channel to train my model, it just need gray images as input, therefore , this can reduce the computational complexity, which for my app, speed is very important.

Answer 1

We don't have any pretrained models that do one channel images. If you're interested in creating a new model, consider adding a new SSD feature extractor.

That being said, you might not see tangible speed ups to your model. Using only one channel only affects the input layer of the network, not the intermediate layers. It might be better to consider using a thinner Mobilenet feature extractor if performance is your highest concern.