search

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).
MIT License
1.73k stars 784 forks source link

RBC detection : No bounding boxes on testing #298

Open sanelec1989 opened 6 years ago

sanelec1989 commented 6 years ago

I have trained the network on only 4 images for larger number of epochs. Following is the snapshot of the losses

Epoch 5137/10000 2018-05-24 17:51:56.121175: I tensorflow/core/kernels/logging_ops.cc:79] Dummy Line [0] 2018-05-24 17:51:56.121259: I tensorflow/core/kernels/logging_ops.cc:79] Loss XY [7.58396491e-05] 2018-05-24 17:51:56.121300: I tensorflow/core/kernels/logging_ops.cc:79] Loss WH [0.0005337451] 2018-05-24 17:51:56.121336: I tensorflow/core/kernels/logging_ops.cc:79] Loss Conf [1.54243735e-05] 2018-05-24 17:51:56.121369: I tensorflow/core/kernels/logging_ops.cc:79] Loss Class [0] 2018-05-24 17:51:56.121404: I tensorflow/core/kernels/logging_ops.cc:79] Total Loss [0.000625009125] 2018-05-24 17:51:56.121435: I tensorflow/core/kernels/logging_ops.cc:79] Current Recall [1] 2018-05-24 17:51:56.121470: I tensorflow/core/kernels/logging_ops.cc:79] Average Recall [0.472331941] 2018-05-24 17:51:56.494739: I tensorflow/core/kernels/logging_ops.cc:79] Dummy Line [0] 2018-05-24 17:51:56.494835: I tensorflow/core/kernels/logging_ops.cc:79] Loss XY [0.0512491763] 2018-05-24 17:51:56.494876: I tensorflow/core/kernels/logging_ops.cc:79] Loss WH [0.180448502] 2018-05-24 17:51:56.494917: I tensorflow/core/kernels/logging_ops.cc:79] Loss Conf [0.00093903509] 2018-05-24 17:51:56.494953: I tensorflow/core/kernels/logging_ops.cc:79] Loss Class [0] 2018-05-24 17:51:56.494991: I tensorflow/core/kernels/logging_ops.cc:79] Total Loss [0.23263672] 2018-05-24 17:51:56.495025: I tensorflow/core/kernels/logging_ops.cc:79] Current Recall [0] 2018-05-24 17:51:56.495061: I tensorflow/core/kernels/logging_ops.cc:79] Average Recall [0.472286] 1/1 [==============================] - 1s 553ms/step - loss: 6.2501e-04 - val_loss: 0.2326

At test time I am using the same training set. However there is no detection. I tried to change the obj_threshold from 0 to 0.5 , but it does not give any good result. I am attaching my training and testing code. Please help me out if I am misssing something

Train_test.zip

saeedalahmari3 commented 6 years ago

Same issue , any idea? what pertained weights was used on the provided notebook?

ArkaJU commented 6 years ago

Did you get the pretrained weights for RBC detection?