readme fix

Edmond Yoo

2018-09-06 595a5ea1f6fed9351862a2708a64d7fc10a7cf6a

 README.md

@@ -10,16 +10,16 @@

First batch of model training is completed, where I used ~40,000 generated images of MTG cards laid out in one of the pre-defined pattern. 

<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_training_set_example_1.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_training_set_example_1.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_training_set_example_1.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_training_set_example_1.jpg" width="360">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_training_set_example_2.jpg" width="360">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_training_set_example_3.jpg" width="360">

After 5000 training epochs, the model got 88% validation accuracy on the generated test set. 

<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_1.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_2.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_3.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_4.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_1.jpg" width="360">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_2.jpg" width="360">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_3.jpg" width="360">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_4.jpg" width="360">

However, there are some blind spots on the model, notably:

@@ -29,8 +29,8 @@

Example of bad detections:

<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_5.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_6.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_7.jpg" width="600">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_5.jpg" width="360">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_6.jpg" width="360">
<img src="https://github.com/hj3yoo/darknet/blob/master/figures/0_detection_result_7.jpg" width="360">

The second and third problems should easily be solved by further augmenting the dataset with random lighting and image skew. I'll have to think more about the first problem, though.