Does Caffe "ImageNet" support .jpg or .bmp images as input?
In ILSVRC2012, I see only .JPEG files. I tried to train models on .jpg and .bmp files (different train for each type), but it doesn't give reasonable results.
Following on to my comment above ...
The Caffe release includes several popular models that you can train with ImageNet data. The input layers of these models assume a consistent format, usually either JPEG or LMDB. An input layer that accepts one format is not useful for an incompatible format, such as BMP.
If you want to accept BMP input instead, you have to specify or write a different input layer. If you want to accept both JPEG and BMP in the same training run, you have to set up separate input processing for the two formats to get them into a compatible form.