SSD & ResNets over black and white pictures?


I often meet vision tasks that could be handled in black and white instead of colors (eg license plate cropping or car shape classification). I believe trainings would be faster if I drop the color channel and run those tasks on black&white matrixes. Is this a common practice, and are there gluon snippets or demos demonstrating resnet and SSDs over black&white pics?

Hi, for prototyping I use Mathematica (which uses MxNet under the hood). For training on greyscale images, I first prepend a layer with 1 x 1 convolution with weights set to 0.3333 and 0 learning rate to network to be trained. This layer transforms color images (jpg/ png input) to grayscale and then “first” real layer operates on a single channel instead of 3. Unfortunately the expected benefits in convergence speed and accuracy hardly noticeable.