• Seanlinx/mtcnn mxnet 4

  • #61: train pnet slow: 1.ssd, 2.imdb, 3.delete all out_grad=True in core\symbol.py 4. 等几分钟就会快起来
  • #57: roc compare 1~2% lower
  

• zuoqing1988/train-mtcnn mxnet-win 借鉴自 Seanlinx/mtcnn 改进可以借鉴

  • #5: DiscROC 准确率。

• AITTSMD/MTCNN-Tensorflow tensorflow 3.5

  • #6: Pnet 准确率问题

• foreverYoungGitHub/MTCNN caffe 3

• CongWeilin/mtcnn-caffe caffe 借鉴自 foreverYoungGitHub/MTCNN imdb 加速可以借鉴

• wujiyang/MTCNN_TRAIN pytorch 2

• mtcnn 训练日志: 主要在 oreverYoungGitHub/MTCNN 上做的尝试

• Anchor Cascade for Efficient Face Detection Anchor overlap

• 生成更多数据： range(more)

• traininghistory: finalize() to dump plot data

• caffe pnet fddb vs mxnet pnet fddb

• default:

  • settings: max-40 [50, 5, 20] 3:1:1
  • train: 984792 (598251 199356 199464)
  • val: 254360 (154425 51511 51536)

• v1: 4:1:1

• v2: 很低 (0.18, 0.9) [50, 10, 20] min(w, h) < 25 or max(w, h) < 30

  • train: 2057290 (1102561 484812 510866)
  • val: 523244 (279235 123882 132353)

  celeba 数据集 生成负样本

• mxnet-mtcnn:

• Sample: 12880 images done, pos: 199475 part: 548912 neg: 812070

• Choose: total 1099474 (pos: 199475 part: 300000 neg: 600000)

• v1 [train] sample: pos = 205458, part = 533325, neg = 766085, total=1504868 [train] filter: pos = 205458, part = 300000, neg = 600000, total=1105458 [val] sample: pos = 52709, part = 137647, neg = 192566, total=382922 [val] filter: pos = 52709, part = 137647, neg = 192566, total=382922

• v2 [train] sample: pos = 195415, part = 540703, neg = 767363, total=1503481 [train] filter: pos = 195415, part = 300000, neg = 600000, total=1095415 [val] sample: pos = 50587, part = 139040, neg = 193038, total=382665 [val] filter: pos = 50587, part = 139040, neg = 193038, total=382665

python tools/fddb/eval.py -s cpnet models/mtcnn/cpnet 27,28,29,30,31,32 3 300000