Vision Transformers and Object Detection

Question 1 of 5

Why is the Vision Transformer (ViT) forced to use relatively large patches for high-resolution images?

Small patches increase token count, causing high complexity in self-attention.Small patches make the model prone to overfitting on fine textures.Positional encodings are only effective for large patch sizes.The patch embedding projection requires a fixed large input size.

Question 2 of 5

Why is Generalized IoU (GIoU) preferred over standard IoU for bounding box regression during training?

GIoU provides gradients even for non-overlapping boxes.GIoU is computationally cheaper than standard IoU.GIoU directly maximizes classification accuracy alongside box localization.GIoU eliminates the need for bipartite matching in object detection.

Question 3 of 5

Why is bipartite matching necessary when training an object detector that outputs a fixed number of prediction slots?

Because ground-truth objects are unordered, requiring optimal assignment to fixed prediction slots.Because the network must predict exactly one bounding box per slot, regardless of object presence.Because the Intersection over Union (IoU) loss is non-differentiable for disjoint boxes.Because the Hungarian algorithm provides the fastest way to compute the training loss.

Question 4 of 5

What enables DETR to avoid using non-maximum suppression (NMS) for duplicate removal?

Decoder self-attention among object queries causes them to specialize and avoid duplicates.The bipartite matching loss penalizes duplicate predictions during training.The Transformer encoder aggregates global context, making duplicate detections inherently unlikely.The feed-forward network predicts a confidence score that is thresholded to eliminate duplicates.

Question 5 of 5

How does deformable attention in Deformable DETR reduce computational complexity compared to standard self-attention?

It samples a few learned points per reference, achieving linear complexity.It restricts attention to non-overlapping windows, similar to the Swin Transformer.It uses a low-rank approximation of the attention matrix to reduce computations.It applies attention only within a single feature map, avoiding multi-scale processing.