WongKinYiu / ScaledYOLOv4

Scaled-YOLOv4: Scaling Cross Stage Partial Network
GNU General Public License v3.0
2.01k stars 570 forks source link

yolov4 tiny #38

Open chingi071 opened 3 years ago

chingi071 commented 3 years ago

您好, 我想請問一下這裡的 yolov4-tiny 跟 darknet 的 yolov4-tiny 架構都一樣,但我看論文有說 backbone改為 OSANet,想請問一下這兩個有什麼差別呢? 另外也想請教一下, yolov4-tiny 裡似乎沒有使用 mosaic 是為什麼呢? 謝謝您。

WongKinYiu commented 3 years ago

您好

yolov4 的 model zoo 裡就有提到 yolov4-tiny 的 backbone 來自 VoVNet. 而 VoVNet 便是由 OSA module 組成. 兩者除了訓練流程外都是相同的.

輕量模型與大型模型最大的不同便是輕量模型一般是參數量不足, 而大型模型則是過參數化. 因此大型模型可以handle rich data augmentation, 而輕量模型難以學習更加泛化的資訊. thundernet paper中有些相關的分析, 可以看看.

chingi071 commented 3 years ago

您好,非常感謝您的回覆!! 我想再請問一點就是訓練圖像支援 .tiff 格式嗎? 還是需要將其轉換為 .jpg 檔呢? 謝謝您。

WongKinYiu commented 3 years ago

基本上opencv可以讀取的格式都支援 如果channel不是3的話需要修改部分程式碼

lh-13 commented 3 years ago

@WongKinYiu
Hello, I found that the yolov4-tiny.cfG file you provide now is the same as the yolov4-tiny.cfg file provided in Darknet around September 2020, except for the number of iterations, why?Wasn't the scaled-yolov4 paper published only in November 2020?I'd like to find out what has happened to the yolov4-tiny structure in Scaled-yolov4

WongKinYiu commented 3 years ago

the only difference is that yolov4-tiny in scaled-yolov4 is trained from scratch.