Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ
Chú thích hình ảnh cho bài kiểm tra thị giác máy tính
Xem lại các lớp hình ảnh, hộp, mặt nạ và điểm chính, quy tắc tắc, sự bất đồng của người chú thích, xem xét nhãn tự động và xuất xứ nhãn.
Câu hỏi 1 của 8
A dataset says an image contains a bicycle but gives no location. Which annotation type is it using?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- A dataset says an image contains a bicycle but gives no location. Which annotation type is it using?
- A separate bounding box for each bicycle
- An instance mask around the bicycle
- A set of keypoints on bicycle parts
- An image-level class label
- A defect detector needs the exact damaged pixels. Which label is most direct?
- Only an image-level presence tag
- An instance or region mask
- A rough box around the entire part
- A category ID without pixel coordinates
- Two annotators disagree on a partly hidden object. What should the team inspect first?
- Whether the written occlusion rule specifies how to label it
- Whether one annotator used a faster interface
- Whether the exported file is smaller
- Whether a model preferred one disputed answer
- Why overlay random exported boxes on their source images?
- Parsing guarantees coordinate correctness
- The overlay proves every class is represented
- A valid file can still contain shifted or wrong geometry
- Visual inspection itself prevents split leakage
- A model proposes masks for annotators to accept. What risk remains?
- Suggestions are always independently correct
- Model errors cannot enter human-reviewed labels
- An auto-labeler eliminates ontology ambiguity
- Its systematic mistakes may be copied into ground truth
- What should a benchmark version record after categories are merged?
- Only the total image bytes
- The changed ontology and which labels used it
- Only the newest model score
- Only the color palette used by annotators
- Why split related video frames by source rather than randomly by image?
- Every video frame has an unrelated source
- Random splitting removes visual similarity
- Near-duplicate frames can leak across train and test
- Video sequences cannot have annotation masks
- What limitation remains even after careful label review?
- The collected images may still miss deployment settings and people
- Accurate labels guarantee global coverage
- A reviewed mask makes collection bias impossible
- Correct boxes eliminate all deployment shift