DETR Transformer Detection Quiz

Answers and explanations appear as you take the quiz.

1. What does DETR remove compared with traditional detectors like Faster R-CNN?
   • Hand-designed components such as anchor boxes and non-maximum suppression
   • The CNN backbone entirely
   • The need for any training data
   • Bounding boxes as an output format
2. Which algorithm does DETR use to match predictions to ground-truth objects during training?
   • k-means clustering
   • Gradient boosting
   • The Hungarian (bipartite) matching algorithm
   • Beam search
3. What are DETR's 'object queries'?
   • Text prompts typed by the user
   • Anchor boxes of fixed sizes
   • Pixels sampled from the image
   • A fixed set of learned embeddings the decoder turns into detections
4. What overall architecture does DETR combine?
   • A recurrent network plus a decision tree
   • A CNN backbone plus a transformer encoder-decoder
   • A pure GAN
   • A k-nearest-neighbor classifier
5. How does DETR's training objective support inference without a separate NMS stage?
   • They make every possible duplicate prediction mathematically impossible
   • The model can only output one object per image
   • One-to-one matching and set losses discourage multiple predictions for the same object
   • A hidden NMS classifier supplies every training target
6. Which was a known weakness of the original DETR?
   • Slow training convergence and weaker small-object detection
   • It could not run on GPUs
   • It required no labeled data
   • It only worked on grayscale images
7. What training label does original DETR assign to an object query that is not matched to a ground-truth object?
   • The largest visible object's class regardless of matching
   • The no-object class
   • A mandatory new ground-truth object invented for the query
   • The image's filename as a class label
8. Which prediction errors does DETR's set loss penalize on matched object pairs?
   • Only the number of letters in the class name
   • Only the image's compression size
   • Only whether the training batch uses a GPU
   • Both the predicted class and bounding-box errors