0

I am working on an object detection project using Yolo-v3 algorithm using python.

Training model on 2 classes: Person and Football. I used a subset of images from the Coco dataset for the Person category and created the football dataset myself.

Training specs are as follows - Images Taken: 3500 * 2 classes = 7000 Batch Size: 64 Epochs: 60 each (unfreezed last layer and later unfreezing full network)

Tech Specs- Nvidia Tesla V100, Ubuntu, 32 gb RAM

Please help me with the following issues: 1. Unable to detect objects with intersecting bounding boxes i.e.,close object detection problem 2. Model is failing to detect both classes in a single picture

The loss functions seem to have stabilised -

Epoch 53/60 707/707 [==============================] - 463s 652ms/step - loss: 8.2134 - val_loss: 8.1215

Epoch 54/60 707/707 [==============================] - 454s 653ms/step - loss: 7.9846 - val_loss: 8.0124

Epoch 55/60 707/707 [==============================] - 461s 653ms/step - loss: 8.1012 - val_loss: 8.1344

Epoch 56/60 707/707 [==============================] - 455s 655ms/step - loss: 8.9814 - val_loss: 8.2155

Epoch 57/60 707/707 [==============================] - 460s 651ms/step - loss: 8.0459 - val_loss: 8.0664

Epoch 58/60 707/707 [==============================] - 465s 657ms/step - loss: 7.9934 - val_loss: 8.1552

Epoch 59/60 707/707 [==============================] - 459s 650ms/step - loss: 7.9364 - val_loss: 8.2638

Epoch 60/60 707/707 [==============================] - 461s 652ms/step - loss: 7.9426 - val_loss: 8.0902

Person and football

Person and person

Thanks in Advance, Michael

Michael
  • 23
  • 4
  • can you try any pretrained weights from coco and see if it detects all ? even in that if it doesnt, in that case, then the problem is your post processing code. – venkata krishnan Apr 26 '20 at 06:33
  • @venkatakrishnan I have checked using pre-trained weights and it is working fine. This is not a problem that I was expecting with using Yolov3. – Michael Apr 27 '20 at 05:45

0 Answers0