mAP score while retraining about yad2k HOT 8 OPEN

I've investigated YAD2K and didn't find calculations for mAP. But you can write down code for mAP in test_yolo.py. I tried this one but mAP is too low, I think I made mistake. Briefly, you have to take score threshold and check if score for that lable is bigger than your threshold you can check IOU and then define is it True Positive or False Positve. Also if there is no bounding box for ground truth label - it should be False Negative. Have to increment score threshold, so do this calculations in loop while Recall begins zero. Then for every class you have Precision and Recall values ( different values for corresponding score thresholds) use it to calculate average precision(https://sanchom.wordpress.com/tag/average-precision/). And after that summarize average precision and divide it by number of classes.

from yad2k.

It will be better to do it using tensorflow. But you can do this just using python, not pretty solution, but quick.

from yad2k.

https://github.com/allanzelener/YAD2K/issues/40
There you can find answer.

from yad2k.

thanx,
i am working on f1 score will share once i test it

from yad2k.

do you have any news about the map score you get? thx

from yad2k.

https://github.com/Pankajchandan/My-Machine-Learning-Rep/blob/master/f1_score.ipynb

this is an implementation for calculating F1 score for multiclass object detection. This can be extended for Map without much hassle.

from yad2k.

@Pankajchandan Did you tested on yad2k? If yes, which accuracy did you get?

from yad2k.

@Firyuza no i have not done that yet. This can be done by calling pred_info(y_true, y_pred) for every image/frame in yad2k. Here y_true is the ground truth in the default format used by yad2k i.e. "class_num x_min y_min x_max y_max" and y_pred is the prediction for that frame/image in the same format. Calling pred_info each time will return a dataframe having classes as indices and "count", "fp", "fn", "tp" and "dnd as columns, You need to keep adding this result somewhere in your code. When all the predictions in you test set are done, you just need to call cal_f1(pred_info) just once to get a data frame having all the classes as indices and "precision", "recall" and "f1 score" as the columns. Here the argument pred_info in cal_f1 is the result of the addition of all the results from the function pred_info.

from yad2k.