nerfren icon indicating copy to clipboard operation
nerfren copied to clipboard

how many phoths are suitable?

Open rabbitchenrui opened this issue 2 years ago • 3 comments

Thanks for your wonderful work ! By your code, I have achieved good results on the real data set of 30 photos. However, when I want to further optmize the effects and input 60 photos, the effects become worse. When I add 560 photos, the new perspective generated is almost invisible. Do you have met similiar problems and how to slove them?

rabbitchenrui avatar Nov 23 '22 03:11 rabbitchenrui

Theoretically, more training images will result in better results. Did you try to train for more epochs/iterations? It's probably because more pixels are used in training if you provide more images, which may require longer training time.

bennyguo avatar Nov 27 '22 09:11 bennyguo

    Thanks for your reply. Acctually, with 40 epoches, it took me a long time (five days)to train 560 photo scenes, but I could hardly see anything in or out of the mirror. I've tried 20, 30, 40, 50, 60 photos these days, and found that the more photos, the worse the effect (all 40 epoches)

陈睿 @.***

 

------------------ 原始邮件 ------------------ 发件人: "bennyguo/nerfren" @.>; 发送时间: 2022年11月27日(星期天) 下午5:47 @.>; @.@.>; 主题: Re: [bennyguo/nerfren] how many phoths are suitable? (Issue #5)

Theoretically, more training images will result in better results. Did you try to train for more epochs/iterations? It's probably because more pixels are used in training if you provide more images, which may require longer training time.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

rabbitchenrui avatar Nov 29 '22 10:11 rabbitchenrui

Did you manual label the reflected regions? Please try training with a few (~5) labeled masks.

bennyguo avatar Dec 08 '22 11:12 bennyguo