Closed xiaoxueshengyao closed 4 years ago
Hi, please see #5 and some of the comments in #3 and #1 for details on improving performance. Basically, you can sweep parameters and the models, but the performance is not very satisfactory. Other methods should be explored to decide LCD classification, e.g. local thresholds of the distance map.
When I used "Inceptin_v1" as the predict model, I can got the score confusion matrix, but Precision-Recall curve is not ideal, especially the precision is always under 0.2, what ever I used v2/v3/v4. That's what I run
python3 cnn_lcd.py --dataset college inception_v1
and the result about precision_recall:Average Precision: 0.16354226114901457 Best MPC-ACC (thresh=0.38): 0.8863359002585536 Precision (thresh=0.38): 0.05770001215400474 Recall (thresh=0.38): 0.8642125134843581
In other way, I think the sim confusion matrix is not suited the precision. How can I improve it. I will be appreciate if you can reply me.