Câu đố về thị giác máy tính

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

1. Phát hiện đối tượng khác với phân loại hình ảnh như thế nào?
   • Classification locates each object with a box, while detection assigns one label to the whole image
   • Classification and detection both return one label for the whole image
   • Detection finds object locations with labels, while classification assigns an image-level label
   • Classification labels pixels, while detection labels regions without locations
2. What does image segmentation produce?
   • One label for the whole image
   • A set of bounding boxes around object instances
   • Labels assigned to pixels or regions
   • Class scores without spatial assignments
3. What can a classifier get wrong even when its image label appears confident?
   • It may rely on a background shortcut rather than the intended object
   • It always locates every object boundary
   • It necessarily checks every relevant condition
   • It proves that the scene is safe
4. Why keep related images from the same scene or recording together in a data split?
   • To balance image labels across training and test sets
   • To prevent near-duplicate scene information from appearing in both partitions
   • To place every image from each category into the test set
   • To increase visual variation in the training examples
5. How can cropping affect an image model?
   • It can remove small objects or useful context before inference
   • It preserves every small detail after compression
   • It automatically reconstructs content outside the cropped area
   • It changes an image classifier into a segmentation model
6. Which evaluation best reflects how a vision model will be used?
   • Only clean studio images from training
   • Realistic camera, lighting, resolution, viewpoint, and environment conditions
   • Only images from one familiar location
   • Only the model’s confidence scores
7. Why should a vision team consider the cost of different mistakes before deployment?
   • False alarms and missed objects have equal consequences in every application
   • The model’s confidence score already incorporates every operational cost
   • Error costs matter only while training, not when the system is used
   • Different errors can have different consequences in the intended application
8. Does a confident visual label prove that a scene is safe?
   • No; confidence is a model score, while safety depends on context, error consequences, and appropriate controls
   • Yes; a sufficiently high confidence score establishes safe use in every setting
   • Yes; strong image-level accuracy proves that each individual prediction is safe
   • Only when the model labels pixels rather than assigning image-level categories