Assign each inspection image a normal or abnormal category. Separately, return the class and rectangular location of each defect in the image.
A pixel-level mask following each defect outline is not required here. Distinguish the whole-image decision from per-object locations.
Choose TWO explanations that match the required outputs.
Image classification assigns image labels, object detection estimates classes and boxes, and segmentation addresses pixel-level regions. Image-level results do not establish localization quality.
Detailed explanation
A box represents an enclosing region, whereas a segmentation mask represents pixel-level areas. They are not identical.
Object classes and bounding boxes match the requested per-object location output.
Category accuracy does not measure displaced boxes or missed defects. Evaluation needs suitable ground truth for the requested location output.
This assigns a label to the image. That label alone does not necessarily provide rectangular locations for each defect.
An image-level label and per-object coordinates are different outputs. Knowing that an image is abnormal does not establish where each defect is.
Check official documentation
Compare the following points with the official documentation.
Compare SageMaker AI documentation for image classification, object detection, and segmentation. Sketch an image label, per-object boxes, and a pixel mask for an invented inspection image, then identify the required outputs.
Official references
- Official documentation 1
- Official documentation 2
- Official documentation 3
- Official documentation 4
- Official documentation 5
What to confirm
Distinguish whole-image categories, per-object rectangles, and pixel masks. Do not report category accuracy as location accuracy. No model inference or accuracy measurement is performed.
Key points
- Image-level labels
- Per-object boxes
- Pixel masks
- Output-specific evaluation
Notes
- Environment: Current AIF-C01 Task 1.1 and related AWS primary-source review using invented examples; no model training or inference is performed.
- Compare SageMaker AI documentation for image classification, object detection, and segmentation. Sketch an image label, per-object boxes, and a pixel mask for an invented inspection image, then identify the required outputs.
- Distinguish whole-image categories, per-object rectangles, and pixel masks. Do not report category accuracy as location accuracy. No model inference or accuracy measurement is performed.
Foundation review
Distinguish the unit of prediction
Deciding whether an entire image is abnormal differs from identifying the class and location of each defect. Check the actual outputs rather than relying only on a model name.
Separate rectangles from pixel regions
An enclosing rectangle is not a mask that follows an object outline. This question requires boxes; a future requirement for pixel-level area or outlines would require reviewing the desired output.
Evaluate against matching ground truth
Correct image labels alone cannot evaluate each defect location. Assess position errors and missed defects against suitable annotations rather than reporting image classification results as localization performance.