Hello, Logan! Thanks for all the work you've put into this repository. I'd like to inquire about some info regarding the training parameters for the evaluation models that you've provided via GDrive: https://drive.google.com/drive/folders/0B9jhaT37ydSyRk9UX0wwX3BpMzQ. Specifically, I'm trying to determine the conv layers' kernel sizes, so I could approximate the minimum region area that is most likely to lose information from the original picture.
My use case involves transferring style for a couple thousand pictures, which are later to be used for an object detection/classification dataset. I've already tried copying the original annotations and using them with the 'augmented' pictures, predictably that resulted in a high number of false positive detections and also some of the crucial fine detail being lost for some of the extracted features (I've used AlexeyAB's YoloV4 for the task). So I need a somewhat reliable way to tell how small a region should be in order to still contain "useful" data for a given model/training picture; I expect an increase of true positives recall figures -> higher detection mAP and thus less negative samples one'd have to add in order to get rid of FPs.
AFAIK, kernel size is num_inputs*num_outputs, but I'm not sure if I should just take num_filters*filter_size or also account for residual blocks (or maybe all of them?) -> filters*filter_size*residual_filter_size (that might be the number of channels, which won't be relevant for what I'd like to accomplish). And whether I should add 1 to count in the preds layers or not. Thank you a lot in advance!
Hello, Logan! Thanks for all the work you've put into this repository. I'd like to inquire about some info regarding the training parameters for the evaluation models that you've provided via GDrive: https://drive.google.com/drive/folders/0B9jhaT37ydSyRk9UX0wwX3BpMzQ. Specifically, I'm trying to determine the conv layers' kernel sizes, so I could approximate the minimum region area that is most likely to lose information from the original picture.
My use case involves transferring style for a couple thousand pictures, which are later to be used for an object detection/classification dataset. I've already tried copying the original annotations and using them with the 'augmented' pictures, predictably that resulted in a high number of false positive detections and also some of the crucial fine detail being lost for some of the extracted features (I've used AlexeyAB's YoloV4 for the task). So I need a somewhat reliable way to tell how small a region should be in order to still contain "useful" data for a given model/training picture; I expect an increase of true positives recall figures -> higher detection mAP and thus less negative samples one'd have to add in order to get rid of FPs.
AFAIK, kernel size is num_inputs*num_outputs, but I'm not sure if I should just take num_filters*filter_size or also account for residual blocks (or maybe all of them?) -> filters*filter_size*residual_filter_size (that might be the number of channels, which won't be relevant for what I'd like to accomplish). And whether I should add 1 to count in the preds layers or not. Thank you a lot in advance!