psyhtest
commented
5 years ago While PIL.Image.LANCZOS may be superior to PIL.Image.BILINEAR, it's incompatible with our current preprocessing (used e.g. in the TFLite program). So we will get different results in 3 cases:
- native TF preprocessing (inside the graph):
PIL.Image.LANCZOSPIL.Image.BILINEAR
I'm wondering if we should try native TF processing outside the graph? In this case, we should get the same precision but with batching enabled.
added the possibility to process batches in the object detection script. However, batch processing request a fixed image dimension, so I had to include a resizing of the processed images. This usually introduces a loss of precision in the network.
I changed the IMAGE_COUNT parameter, and it is now divided into 2 variables, BATCH_COUNT and BATCH_SIZE, with the number of processed images being the product of the two.
Other introduced parameter is the ENABLE_BATCH, that is set to false as default, to maintain the previous program ignoring all the resizing and modification for the batch run.
Moreover the resize is done giving 2 parameters, height and width, to support the possibility of resizing to non-squared images. to resize is used the LANCZOS algorithm of the pillow library. The API says that is an high-quality downsizing algorithm.