[BUG] Pictures in portrait mode are not rendered correctly. Maybe causing an issue with training
ultratoto14 opened this issue · 11 comments
I went from original double-take to yours and found out that all my pictures coming from my phone (Samsung Galaxy Note 20 Ultra) in portrait mode are displayed in landscape mode in the UI.
I thought it was not a problem but when I was trying CodeProject.AI, the logs shows no face detected on these images, Just added one picture taken in landscape mode and CodeProject.AI detected the faces.
Version of Double Take
1.13.11.3
Expected behavior
Portrait photo should be sent/viewed as portrait
Hardware
- amd64 + GPU
- Qnap NAS OS
- Chrome
- skrashevich/double-take:latest
Do you upload photos directly from your phone, or do you transfer them to PC first?
Hi, I uploaded from phone directly, using google chrome.
please, provide an example of problem photo file
thx. I was able to reproduce this issue. fixed in current latest beta build. will be included in next release
Hi @skrashevich , just tested the latest beta and it seems that it works most of the time.
When I tried to retrain all the files sent from my phone, it's really slow to process and most of them received a 413 from compreface:
172.29.36.1 - - [25/Sep/2023:06:59:40 -0700] "POST /api/v1/recognition/faces?subject=emile HTTP/1.1" 201 91 "-" "axios/1.5.0" 2023/09/25 07:00:29 [error] 106#106: *651862 client intended to send too large body: 5338079 bytes, client: 172.29.36.1, server: ui, request: "POST /api/v1/recognition/faces?subject=emile HTTP/1.1", host: "192.168.15.4:8000" 172.29.36.1 - - [25/Sep/2023:07:00:30 -0700] "POST /api/v1/recognition/faces?subject=emile HTTP/1.1" 413 192 "-" "axios/1.5.0" 2023/09/25 07:03:16 [error] 106#106: *651863 client intended to send too large body: 5206261 bytes, client: 172.29.36.1, server: ui, request: "POST /api/v1/recognition/faces?subject=jade HTTP/1.1", host: "192.168.15.4:8000" 172.29.36.1 - - [25/Sep/2023:07:03:16 -0700] "POST /api/v1/recognition/faces?subject=jade HTTP/1.1" 413 192 "-" "axios/1.5.0" 2023/09/25 07:03:16 [error] 106#106: *651864 client intended to send too large body: 5139979 bytes, client: 172.29.36.1, server: ui, request: "POST /api/v1/recognition/faces?subject=jade HTTP/1.1", host: "192.168.15.4:8000" 172.29.36.1 - - [25/Sep/2023:07:03:16 -0700] "POST /api/v1/recognition/faces?subject=jade HTTP/1.1" 413 192 "-" "axios/1.5.0" 2023/09/25 07:03:16 [error] 106#106: *651865 client intended to send too large body: 5130935 bytes, client: 172.29.36.1, server: ui, request: "POST /api/v1/recognition/faces?subject=jade HTTP/1.1", host: "192.168.15.4:8000" 172.29.36.1 - - [25/Sep/2023:07:03:16 -0700] "POST /api/v1/recognition/faces?subject=jade HTTP/1.1" 413 192 "-" "axios/1.5.0"
As it seems that you changed the image orientation, the picture data size became too big.
it's not a double-take problem. look like you have nginx in front of compreface?
I use the single image of compreface:
compreface:
container_name: compreface
restart: always
image: exadel/compreface:1.0.0-arcface-r100-gpu
Effectively, when I go inside the compreface docker container, it seems to run an nginx instance.
The problem remains the same, the original picture go through double-take to compreface with the orientation problem. With you fix, some pictures are going through with the right orientation but in some case, the rotated picture is too large for compreface because it's way bigger than the original picture.
this is limitation in compreface single-image: https://github.com/exadel-inc/CompreFace/blob/4d826e53bdbe82c2a2ed2229aa27785cdf732247/custom-builds/Single-Docker-File/nginx.conf#L19
for increase this limit you can, for example, download nginx.conf to you host, change this property and mount it to compreface container (docker run .... -v path_to_modified_nginxconf:/etc/nginx/conf.d/nginx.conf
)
anyway, this is compreface issue, not double-take
thx. I was able to reproduce this issue. fixed in current latest beta build. will be included in next release
@skrashevich I have the same issue all of a sudden (auto-rotation). Was wondering how this happened, never had an issue with it before. Did we include some new dependancy or something?
Thanks for fixing.
How do i install the beta from the add-on store ? Anyway of doing this ? Or can we push the new release?
fixed in v1.13.11.4 release