Upgradations for Defect Detection
We have 4 Layers (4 Defect Types):
F (PT);
I (Pp1, Tp1);
ITO (Pp1, Tp1);
M2 (LD, Pp1, Tp1)
Embedded Files
1) Issues:
1.1 Miss between TP1 and PT:
1.2 Nearby patches vs. distant patches of an Image
In general, talking about images, a lot of work is done on looking at nearby image patches versus distant patches, so most of the CPC v1 and CPC v2 methods are really exploiting this property of images. So image patches that are close are called as positives and image patches that are further apart are translated as negatives, and the goal is to minimize the contrastive loss using this definition of positives and negatives.
1.3 Patches of an image vs. patches of other images
The more popular or performant way of doing this is to look at patches coming from an image and contrast them with patches coming from a different image. This forms the basis of a lot of popular methods like instance discrimination, MoCo, PIRL, SimCLR. The idea is basically what’s shown in the image. To go into more details, what these methods do is to extract completely random patches from an image. These patches can be overlapping, they can actually become contained within one another or they can be completely falling apart and then apply some data augmentation. In this case, say, a colour chattering or removing the colour or so on. And then these two patches are defined to be positive examples. Another patch is extracted from a different image. And this is again a random patch and that basically becomes your negatives. And a lot of these methods will extract a lot of negative patches and then they will basically perform contrastive learning. So there are relating two positive samples, but there are a lot of negative samples to do contrastive learning against.
2) Ideas from Authors:
2.1 Using YOLO as a multi-label classifier: how? [Link]
We have done this in 2021-01-14, but not yet have the plan to release the function. (hint: GAP and BCE)
And if your mask and glasses has keypoints annotation, you could combine this with our code. (hint: Wing loss)
2.2 Reduce False Postives [Link]
We have done this in 2021-01-14, but not yet have the plan to release the function. (hint: GAP and BCE)
And if your mask and glasses has keypoints annotation, you could combine this with our code. (hint: Wing loss)
2.3 Change Augmentations [Link]
Set mosaic=0 in your hyp.yaml to disable mosaic augmentation.
Add new arguments in hye.yaml and add corresponding code in dataset.py.
3) Parts to Improve:
3.0 Augmentations
Mosaic augmentations
3.2 Neck
3.4 Loss:
Focal loss
4) Other Thinsgs to Improve:
4.1 Recommendation about the Custom Dataset:
Images per class. ≥1.5k images per class
Instances per class. ≥10k instances (labeled objects) per class total
Image variety. Must be representative of deployed environment. For real-world use cases we recommend images from different times of day, different seasons, different weather, different lighting, different angles, different sources (scraped online, collected locally, different cameras) etc.
Label consistency. All instances of all classes in all images must be labelled. Partial labelling will not work.
Label accuracy. Labels must closely enclose each object. No space should exist between an object and it's bounding box. No objects should be missing a label.
Background images. Background images are images with no objects that are added to a dataset to reduce False Positives (FP). We recommend about 0-10% background images to help reduce FPs (COCO has 1000 background images for reference, 1% of the total).
Page updated
Google Sites
Report abuse