exadel-inc / exadel-inc/CompreFace
Inconsistent detection based on face orientation
- Dominant language
- Java
- Stars
- 8.3k
- Forks
- 1.2k
- PR merge metrics
- No merged PRs in 30d
Description
**Describe the bug**
I have a 1080p Wyze Cam v3 that is pointed down at a baby bed and I've been testing to try and determine which model is best for detecting my face. At first I thought all of the models weren't working. I am testing with submitting a picture of my face to Compreface leaning over the bed to be in view of where a baby will be. I had a thought that maybe it was the orientation.
In the regular photo orientation my face is upside down, then I copied the photo, rotated it 90, 180, and 270 degrees.
In normal orientation (0 degrees) where my face is oriented upside down, it cannot detect any faces.
In 90 and 180 degrees it correctly detects my face.
And in 270 degrees it doesn't detect my face.
Is this behavior normal / to be expected? If so I can try to predict orientation and maybe rotate the photo before passing it to compreface?
**To Reproduce**
See above description.
I tried originally with the default model FaceNet and it didn't detect my face in the 0 degrees orientation, so I moved on to SubCenter-ArcFace-r100 and the testing above is done with that custom build.
**Expected behavior**
My face to be detected in all orientations of the photo.
**Screenshots**
If applicable, add screenshots to help explain your problem.
**Desktop (please complete the following information):**
- OS: Ubuntu LTS 22.04
- Browser [e.g. chrome, safari] Edge
- Version [e.g. 22] Version 112.0.1722.48 (Official build) (64-bit)
**Logs**
Run those commands and attach result to the ticket:
`docker ps`
```
# docker ps
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES
d11cd0a1526c exadel/compreface-fe:1.1.0 "/docker-entrypoint.…" 6 hours ago Up 6 hours 0.0.0.0:8000->80/tcp compreface-ui
9598f49979f0 exadel/compreface-admin:1.1.0 "sh -c 'java $ADMIN_…" 6 hours ago Up 6 hours compreface-admin
d220f1373b4b exadel/compreface-api:1.1.0 "sh -c 'java $API_JA…" 6 hours ago Up 6 hours compreface-api
ec7be1d3019e exadel/compreface-postgres-db:1.1.0 "docker-entrypoint.s…" 6 hours ago Up 6 hours 5432/tcp compreface-postgres-db
e836ba1aae3a exadel/compreface-core:1.1.0-arcface-r100 "uwsgi --ini uwsgi.i…" 6 hours ago Up 6 hours 3000/tcp compreface-core
1b637fee06de robmarkcole/deepstack-ui:latest "streamlit run deeps…" 8 hours ago Up 8 hours 0.0.0.0:8501->8501/tcp deepstack-ui
5e1196338c0c deepquestai/deepstack:latest "/app/server/server" 8 hours ago Up 8 hours 0.0.0.0:81->5000/tcp deepstack
f4e5826d50c7 portainer/agent:2.17.1 "./agent" 28 hours ago Up 28 hours 0.0.0.0:9001->9001/tcp portainer_agent
```
`docker-compose logs`
Not sure this is relevant; if needed I can sanitize and provide.
**Additional context**
Thank you -- any help / guidance / ideas are helpful!
Contributor guide
Research direction
Reproduce the report by submitting the same face image to CompreFace at 0°, 90°, 180°, and 270° orientations, using FaceNet and SubCenter-ArcFace-r100. Start with the running Docker services and the selected model; done means documenting whether this behavior is expected or identifying a concrete component and test to change.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- docker, docker-compose
- Domain
- computer-vision, machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 30/100