Training Custom Dataset

[RJVisee44]

Hi all,

I followed all the steps to train a custom dataset and mostly has gone okay. Just wanted to confirm a few things. Note I'm training a dataset that distinguishes between left and right hand. Many training/test images have no hands in the image (i.e. no true positives).

1. labelmap.prototxt requires "background" label? Do I need to do anything special with the annotation files that have no hands within the image? Or will no ground truth bounding boxes be classified as "background" (none of the above)?

item { name: "none_of_the_above" label: 0 display_name: "background" } item { name: "L" label: 1 display_name: "L" } item { name: "R" label: 2 display_name: "R" }

2. Test images are not seen/used during training correct? i.e. if I run the final trained model on the test set after training, it will be evaluating it on data it has not yet seen?

3. Are there any parameters that would increase performance when distinguishing between left/right? I know YOLO has a "flip" parameter which greatly increased performance.

Thanks guys.

[FSet89]

Hi @RyanCodes44, did you find an answer to this question? I am interested too. Thanks

[RJVisee44]

@FSet89

1. I'm not really sure but I don't remember changing anything and it worked well. So I believe the answer to be yes.

2. I don't believe they are.

3. there is a "mirror" parameter. I believe it is in the train.prototxt. Just set it to true.