Eric Liu

Results 85 comments of Eric Liu

Not yet , but , however , As I see it is compatible with this project , your can try it . Contribution is welcome The link was invalid ,...

Thanks for your interesting , but currently , I still try to find the strange problem and want solve it. 1. GIOU : mAP will be increase when training coco...

The default value of use_logic_gradient_ was "false" , https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/layers/yolov3_layer.cpp#L441-#L471

@unclejokerjoker Because [darknet](https://github.com/pjreddie/darknet) didn't use logic_gradient , it only use it in old version "yolov2"

1. 我的mobilenetv3 是從這個[專案](https://github.com/d-li14/mobilenetv3.pytorch)轉過來的,轉完後的model放在[這](https://github.com/eric612/MobilenetV3-Caffe),有部分修改但應該沒有裁減,但因為pytorch 的模型是非官方的,或許會跟paper有些出入,比如說後面的fc ,但基本不影響backbone。 2. 一般我在做小型網路的偵測時,不太用到1/8 scale,雖然可以增加不少小物件偵測率,但upsample會耗費太多計算,但如果不用upsample做特徵融合,1/8 scale的layer太淺層反而對精準度沒有幫助

當然會影響檢測效果,但當初作者用意是在於 multi-scale的訓練,可以在inferernce時根據狀況調整輸入解析度,調控精準度與速度,有這方面的疑問,你可以用416訓練一次,再用352訓練一次測試mAP,就知道差異,一般voc2007 可能會差到1%左右

1. Training 跟 test的anchor 要一致 2. 試者把會random的參數範圍調小,如jitter , expand ratio ... 等 。另外也可以把batch size加大。

沒有採用預訓練的話,建議是先訓練imagenet 1000類網路,訓練完之後再訓練coco,沒有預訓練即使能訓練得起來,mAP也會很低

你看看是不是這個問題 input layer layer { name: "data" type: "Input" top: "data" input_param { shape: { dim: 1 dim: 3 dim: 227 dim: 227 } } } https://github.com/Tencent/ncnn/blob/8d984f105b1b8f816654ffa7719e96fbd5885971/docs/how-to-use-and-FAQ/use-ncnn-with-alexnet.md