Revision history [back]

I have two datasets: first one is 340 images size of 440x200, on every image is objects of size between 9x10 px and 25x15 px. And I have a negative set about 570 images size of 440x200. I run training with these params: -numPos 250 -numNeg 22000 -w 10 -h 10 --numStages 24 --minHitRate 0.995 -maxFalseAlarm -maxDepth 20 -maxWeakCount 600 -mode ALL

I created my set from previosly marked up images.

I also tried -numPos 320 and 300. In first time it end with error "Unsufficient count of positive samples" on 4 stage(-numPos 320). I rerun training on same cascade with -numPos 300. It gives me another one stage and fall with same error. In third time I choose -numPos 250 and it fall on 2 Stage and FalseAlarmRate in 1 Stage was 0! But in second stage 0.028.

So I will be glad to hear any advice about that, but my question is: It is impossible to train cascade with these small set of positives samples? But it give me very small FalseAlarmRate even on first stages. So it can be a good classifier?

I have two datasets: first one is dataset: 340 images size of 440x200, on every image is objects of size between 9x10 px and 25x15 px. And I have a negative set about 570 images size of 440x200. I run training with these params: -numPos 250 -numNeg 22000 -w 10 -h 10 --numStages 24 --minHitRate 0.995 -maxFalseAlarm -maxDepth 20 -maxWeakCount 600 -mode ALL

I created my set from previosly marked up images.

I also tried -numPos 320 and 300. In first time it end with error "Unsufficient count of positive samples" on 4 stage(-numPos 320). I rerun training on same cascade with -numPos 300. It gives me another one stage and fall with same error. In third time I choose -numPos 250 and it fall on 2 Stage and FalseAlarmRate in 1 Stage was 0! But in second stage 0.028.

So I will be glad to hear any advice about that, but my question is: It is impossible to train cascade with these small set of positives samples? But it give me very small FalseAlarmRate even on first stages. So it can be a good classifier?

I have dataset: 340 images size of 440x200, on every image is objects of size between 9x10 px and 25x15 px. And I have a negative set about 570 images size of 440x200. I run training with these params: -numPos 250 -numNeg 22000 -w 10 -h 10 --numStages 24 --minHitRate 0.995 -maxFalseAlarm -maxDepth 20 -maxWeakCount 600 -mode ALL

I created my set from previosly marked up images.

I also tried -numPos 320 and 300. In first time it end with error "Unsufficient count of positive samples" on 4 stage(-numPos 320). I rerun training on same cascade with -numPos 300. It gives me another one stage and fall with same error. In third time I choose -numPos 250 and it fall on 2 Stage and FalseAlarmRate in 1 Stage was 0! But in second stage 0.028.

So I will be glad to hear any advice about that, but my question is: It is impossible to train cascade with these small set of positives samples? But it give me very small FalseAlarmRate even on first stages. So it can be a good classifier?

I have dataset: 340 images size of 440x200, on every image is objects of size between 9x10 px and 25x15 px. And I have a negative set about 570 images size of 440x200. I run training with these params: -numPos 250 -numNeg 22000 -w 10 -h 10 --numStages 24 --minHitRate 0.995 -maxFalseAlarm -maxDepth 20 -maxWeakCount 600 -mode ALL

I created my set from previosly marked up images.

I also tried -numPos 320 and 300. In first time it end with error "Unsufficient count of positive samples" on 4 stage(-numPos 320). I rerun training on same cascade with -numPos 300. It gives me another one stage and fall with same error. In third time I choose -numPos 250 and it fall on 2 Stage and FalseAlarmRate in 1 Stage was 0! But in second stage 0.028.

So I will be glad to hear any advice about that, but my question is: It is impossible to train cascade with these small set of positives samples? But it give me very small FalseAlarmRate even on first stages. So it can be a good classifier?

I have dataset: 340 images size of 440x200, on every image is objects of size between 9x10 px and 25x15 px. And I have a negative set about 570 images size of 440x200. I run training with these params: -numPos 250 -numNeg 22000 -w 10 -h 10 --numStages 24 --minHitRate 0.995 -maxFalseAlarm -maxDepth -0.5 maxDepth 20 -maxWeakCount 600 -mode ALL

I created my set from previosly marked up images.

I also tried -numPos 320 and 300. In first time it end with error "Unsufficient count of positive samples" on 4 stage(-numPos 320). I rerun training on same cascade with -numPos 300. It gives me another one stage and fall with same error. In third time I choose -numPos 250 and it fall on 2 Stage and FalseAlarmRate in 1 Stage was 0! But in second stage 0.028.

So I will be glad to hear any advice about that, but my question is: It is impossible to train cascade with these small set of positives samples? But it give me very small FalseAlarmRate even on first stages. So it can be a good classifier?

It is no face detection problem.

I have dataset: 340 images size of 440x200, on every image is objects of size between 9x10 px and 25x15 px. And I have a negative set about 570 images size of 440x200. I run training with these params: -numPos 250 -numNeg 22000 -w 10 -h 10 --numStages 24 --minHitRate 0.995 -maxFalseAlarm -maxFalseAlarmRate -0.5 maxDepth 20 -maxWeakCount 600 -mode ALL

I created my set from previosly marked up images.

I also tried -numPos 320 and 300. In first time it end with error "Unsufficient count of positive samples" on 4 stage(-numPos 320). I rerun training on same cascade with -numPos 300. It gives me another one stage and fall with same error. In third time I choose -numPos 250 and it fall on 2 Stage and FalseAlarmRate in 1 Stage was 0! But in second stage 0.028.

So I will be glad to hear any advice about that, but my question is: It is impossible to train cascade with these small set of positives samples? But it give me very small FalseAlarmRate even on first stages. So it can be a good classifier?

It is no face detection problem.

I have dataset: 340 images size of 440x200, on every image is objects of size between 9x10 px and 25x15 px. And I have a negative set about 570 images size of 440x200. I run training with these params: -numPos 250 -numNeg 22000 -w 10 -h 10 --numStages -numStages 24 --minHitRate -minHitRate 0.995 -maxFalseAlarmRate -0.5 maxDepth 20 -maxWeakCount 600 -mode ALL

I created my set from previosly marked up images.

I also tried -numPos 320 and 300. In first time it end with error "Unsufficient count of positive samples" on 4 stage(-numPos 320). I rerun training on same cascade with -numPos 300. It gives me another one stage and fall with same error. In third time I choose -numPos 250 and it fall on 2 Stage and FalseAlarmRate in 1 Stage was 0! But in second stage 0.028.

So I will be glad to hear any advice about that, but my question is: It is impossible to train cascade with these small set of positives samples? But it give me very small FalseAlarmRate even on first stages. So it can be a good classifier?

It is no face detection problem.