search

prabindh / darknet-cpp-windows

Visual studio project files for Darknet-cpp inference
35 stars 18 forks source link

No Detection with Custom Models #5

Closed Optimus1072 closed 6 years ago

Optimus1072 commented 6 years ago

Hi, I trained custom models and when I tested those models with original darknet C code everything seems to work fine. I am getting proper detection results. But when I use same config/weights with this C++ extension I am getting 0 detections.

Note : This C++ wrapper is working properly with the pretrained YOLO(both tiny and full) models, the problem persists only when trained on custom dataset. I even tried keeping the config file same and then training on custom dataset, even then I am getting zero detections.

Optimus1072 commented 6 years ago

Further, I would like to add that I haven't trained using the C++ API(if there is an option in it). Training was done using the original C code

prabindh commented 6 years ago

Could you confirm - if the same code-base is used for training (on Linux, as Windows training is not supported yet) ?

Optimus1072 commented 6 years ago

No I have used the original code from darknet for training and C++ wrapper for testing. Yes everything is on Linux. Do you thing I should train also with this wrapper? Sorry for the late reply.

prabindh commented 6 years ago

If you had trained with the latest darknet (original code), it could be an issue with the weight versions. Could you please train with the same code-base, then use the C++ wrapper ? For reference on this, see https://github.com/prabindh/darknet/issues/54

prabindh commented 6 years ago

The code-base is now validated with Yolov3 as well. Closing.