Crop through hairline throws a no face error #155
-
|
Not sure if this is an issue for all users or just me (or just a quirk of face detection), so I'm going to hold off filing a bug report for it. If the target image contains a crop through the hairline near the face, I get a "No faces detected" message. I tested this a few times, below is an example. This image works fine, and the face is detected: This image fails face detection with 60px cropped from the top of the image: Is there a setting I'm missing? Anyone else seeing this happen? Just a quirk to live with? I've seen some talk about det-size, but with the new GUI, there is no option to tweak this (if it can even be changed; or if it would even help). |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments 1 reply
-
|
Had the same problem - top was missing almost the same amount as your example (source face). For source face, adding a black border around the face worked for me. For target face, there actually is an option to tweak it. Just uncheck the "Use default Det-Size" checkbox. Still no guarantee that every frame will be swapped and the result might miss some face details. For detection using "Use Face from this Frame", the face needs to be "complete" as far as I know. But haven't tried with unchecked "use default det-size", always found a frame with complete face for detection. source face with your example (save to see properly, since github's background is too dark to see): |
Beta Was this translation helpful? Give feedback.
-
|
Just tried your example with det size unchecked and it is still not detected. Nice find, this seems to be a limitation of the insightface model 🤷♂️ |
Beta Was this translation helpful? Give feedback.
-
|
Yes I think so, that's why I didn't open an issue. At least with source faces, my "border-workaround" has done the trick twice for me so far. |
Beta Was this translation helpful? Give feedback.
-
|
Hi @lysxelapsed ，good trick. Could you plz share your "border-workaround" trick in more detail? If there are various kinds of images, do you have any insights on how to process them in one function to improve face detection accuracy? |
Beta Was this translation helpful? Give feedback.
Had the same problem - top was missing almost the same amount as your example (source face). For source face, adding a black border around the face worked for me. For target face, there actually is an option to tweak it. Just uncheck the "Use default Det-Size" checkbox. Still no guarantee that every frame will be swapped and the result might miss some face details. For detection using "Use Face from this Frame", the face needs to be "complete" as far as I know. But haven't tried with unchecked "use default det-size", always found a frame with complete face for detection.
source face with your example (save to see properly, since github's background is too dark to see):