how to get higher quality or accuracy of own made Haar cascade - opencv

Question

Could anyone please just roughly tell me what is minimal hit rate, false alarm rate and how do I set width and height for the training purpose. I have already read through documentation on cv2 and also google some of it but in fact it didnt help me much. I have already done my first cascade but it didnt work well and quite horrible. Please just roughly tell me what happen if i change the value of these rate. Im using GUI haar cascade trainer on window. Thanks in advance.

Answer 1

not an answer, but a hint:
e.g. if you have, after stage 0, this result:
NEG count : acceptanceRatio 40000 : 1
Precalculation time: 24.031
+----+---------+---------+
| N | HR | FA |
+----+---------+---------+
| 1| 1| 1|
+----+---------+---------+
| 2| 1| 1|
+----+---------+---------+
| 3| 0.995179| 0.0838|
+----+---------+---------+
END>
Training until now has taken 0 days 0 hours 1 minutes 57 seconds.
and for stage 1 you get:
NEG count : acceptanceRatio 40000 : 0.124695
Precalculation time: 19.241
+----+---------+---------+
| N | HR | FA |
+----+---------+---------+
| 1| 1| 1|
+----+---------+---------+
| 2| 1| 1|
+----+---------+---------+
| 3| 0.999077| 0.142975|
+----+---------+---------+
END>
Training until now has taken 0 days 0 hours 4 minutes 9 seconds.
this means the classifier is quite simple in the beginning. FA of 0.0838 vs. acceptanceRatio 40000 : 0.124695 means, that generalization is ok, so far (0.0838 is close to 0.124695) but there is some gap, so negative samples might be diverse enough. In stage 2, NEG count : acceptanceRatio 40000 : 0.034703 shows, that generalization is still on a good way. 0.124695 * 0.142975 = 0.01782826762 though .
From my experience, the acceptance ratio is one of the most important things to observe during training, to show you the quality of your training data.