You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The OcvYoloDetection component acquires a Triton client lock before attempting to resize an input frame to the YOLO model input dimension size (608 x 608). The resize attempt can throw an exception if the frame dimensions are skewed heavily towards width or height - for example, an image that is 2000 px wide by 1 px tall.
Specifically, the issue will occur when max(w, h) / min(w, h) >= 608 * 2. The OpenCV resize function rounds up fractional parts of a pixel that are more than 0.5.
Throwing the exception results in a deadlock since the top-most exception handler waits until add client locks are freed before terminating the job. Normally, the locks are not freed until after the Triton server sends a response.
The text was updated successfully, but these errors were encountered:
jrobble
changed the title
OcvYoloDetection component deadlocks with odd frame sizes when using Triton
OcvYoloDetection component deadlocks with strange frame sizes when using Triton
Jun 28, 2023
The OcvYoloDetection component acquires a Triton client lock before attempting to resize an input frame to the YOLO model input dimension size (608 x 608). The resize attempt can throw an exception if the frame dimensions are skewed heavily towards width or height - for example, an image that is 2000 px wide by 1 px tall.
Specifically, the issue will occur when
max(w, h) / min(w, h) >= 608 * 2
. The OpenCV resize function rounds up fractional parts of a pixel that are more than 0.5.Throwing the exception results in a deadlock since the top-most exception handler waits until add client locks are freed before terminating the job. Normally, the locks are not freed until after the Triton server sends a response.
The text was updated successfully, but these errors were encountered: