I would like a option to analyze retail shelf photographs, detect the shelf rail/edge angle relative to a horizontal line, and return “Not Detectable” if no legitimate shelf line is discovered. If the detected deviation is greater than 15 levels from horizontal, it must be flagged as misaligned.
The reference photographs present shelf rails, worth strips, and product rows that may very well be used as horizontal cues.
What I would like recommendation on
-
Essentially the most dependable option to detect shelf horizontal strains in a cell app.
-
How you can calculate the angle of detected strains from horizontal.
-
How you can determine when the shelf is not detectable.
-
Greatest practices for dealing with blurry, occluded, or low-contrast shelf photos.
I tried to make use of OpenCV with the Probabilistic Hough Line Remodel and Apple’s VNDetectHorizonRequest.**
Nonetheless, the outcomes haven’t been correct or dependable.** The primary points I am going through:
-
Hough strains picks up too many non-shelf strains (product edges, label textual content, reflections)
-
Outcomes differ considerably with small modifications in lighting or product density
-
Can’t confidently decide when there are actually "no detectable" shelf strains vs. simply noisy detection
-
If Cabinets are actually angled, it offers incorrect outcomes
What I am in search of:
-
Is there a greater pre-processing pipeline (morphological ops, masking, and so forth.) to isolate shelf rails particularly earlier than operating Hough?
-
Is there a scoring/confidence threshold technique to determine when to return
"Not Detectable"vs. a low-confidence angle? -
Are there various approaches on iOS (CoreML, ARKit planes, CoreMotion) that will give extra constant outcomes for this particular use case?
I need customers to keep away from importing photos like these (can not see merchandise clearly from angles, cabinets not detectable):
[!’Example 1′]
