Unit 5: Image Segmentation and Advanced Vision Architectures - Practice Quiz

1 What is the main goal of semantic segmentation?

2 How does semantic segmentation usually treat two objects of the same class?

3 What does instance segmentation produce for each detected object?

4 Which feature distinguishes instance segmentation from semantic segmentation?

5 Why is the UNet architecture named "U-Net"?

6 What is the main purpose of skip connections in UNet?

7 What is the primary role of the decoder path in UNet?

8 Which type of convolution is widely used in DeepLab to capture wider context?

9 What is the purpose of Atrous Spatial Pyramid Pooling in DeepLab?

10 Mask R-CNN extends Faster R-CNN by adding which output branch?

11 What does the mask branch of Mask R-CNN predict?

12 What does self-attention allow each input token to do?

13 Which three components are commonly used to compute self-attention?

14 Why does multi-head attention use several attention heads?

15 How are the outputs of attention heads commonly combined?

16 Why is positional encoding added to Transformer inputs?

17 Which mechanism is central to Transformer-based models?

18 How does a Vision Transformer typically prepare an image for processing?

19 In a standard ViT classifier, what is the main role of the class token?

20 What type of attention is a key feature of the Swin Transformer?

21 A semantic segmentation model processes a street image containing three cars and two pedestrians. How should the output represent these objects?

22 A segmentation model predicts the correct class for most large background regions but performs poorly on small foreground objects. Which metric would better reveal this weakness than overall pixel accuracy?

23 An application must count individual apples in an image even when several apples touch or overlap. Which output is most appropriate?

24 Two predicted masks correspond to the same ground-truth object. Their intersection over union values are and , and the matching threshold is . What is the usual evaluation outcome?

25 In UNet, why are encoder feature maps concatenated with decoder feature maps through skip connections?

26 A UNet model correctly identifies organs but produces coarse boundaries. Which modification most directly increases the decoder's output resolution?

27 A DeepLab model uses an atrous convolution with kernel size and dilation rate . What is its effective kernel size along one spatial dimension?

28 Why does Atrous Spatial Pyramid Pooling use several atrous convolution rates in parallel?

29 In Mask R-CNN, which component preserves spatial alignment when extracting features for a proposed region?

30 A Mask R-CNN system detects an object with the correct class and bounding box, but its predicted mask is inaccurate. Which branch most directly requires improvement?

31 In scaled dot-product self-attention, why is divided by before applying softmax?

32 Suppose one attention row has pre-softmax scores . What attention distribution is produced for the three tokens?

33 What practical benefit does multi-head attention provide compared with a single attention head of similar total size?

34 A transformer has model dimension and uses attention heads with equal dimensions. What is the usual query dimension of each head?

35 A vision transformer receives the same set of patch embeddings in two different spatial orders but uses no positional encoding. What limitation is most likely?

36 A model trained on image patches is fine-tuned using a patch grid. What is a common way to adapt its learned positional embeddings?

37 Why can full self-attention become expensive for high-resolution feature maps containing tokens?

38 A image is divided into non-overlapping patches for a ViT. Excluding the class token, how many patch tokens are created?

39 In a standard ViT classifier, what is the primary role of the class token after the transformer encoder?

40 Why does a Swin Transformer alternate between regular window attention and shifted-window attention?

41 A semantic segmentation model predicts a class label for every pixel, but two adjacent cars touching each other are merged into one connected region. Which change directly addresses the limitation while preserving per-pixel classification?

42 An instance segmentation system achieves high mask IoU on isolated objects but frequently assigns overlapping pixels to multiple objects in crowded scenes. Which evaluation behavior is most likely to expose this failure?

43 In a UNet, skip connections concatenate encoder features with upsampled decoder features at matching resolutions. What is the principal effect when the encoder features retain fine spatial detail but have weak semantic context?

44 A UNet is trained on a dataset where foreground pixels occupy only 2% of each image. The model predicts background everywhere and obtains high pixel accuracy. Which loss modification most directly counters this failure?

45 DeepLab uses atrous convolution with dilation rate . For a kernel, what is the effective kernel size, and what trade-off does increasing introduce?

46 A DeepLab model uses ASPP branches with dilation rates that share a common factor. The output shows repeated checkerboard-like gaps in textured regions. Which explanation is most appropriate?

47 In DeepLab, output stride is changed from 16 to 8 while the input and decoder remain otherwise unchanged. Which consequence is most expected?

48 Why does Mask R-CNN use RoIAlign rather than RoIPool for predicting object masks?

49 Mask R-CNN predicts a binary mask for each detected RoI using a class-specific mask branch. During inference, which operation converts that mask into the final class-conditioned instance mask?

50 A small object is correctly localized by the bounding-box branch but receives a poor mask because its RoI feature is too coarse. Which architectural adjustment most directly improves this issue?

51 For self-attention with queries , keys , and values , the output is . What problem does the factor primarily address?

52 In a vision transformer, a global self-attention layer processes image tokens with embedding dimension . Which asymptotic cost is the main bottleneck as image resolution increases?

53 Why can multi-head attention represent relationships that a single attention head with the same total embedding dimension may miss?

54 A transformer receives image patch embeddings without any positional information. Two images contain identical patch content but with patches permuted spatially. What behavior should be expected from the transformer encoder?

55 A model trained with learned absolute positional embeddings at one fixed image resolution is evaluated at a larger resolution. Which adaptation is commonly required?

56 A transformer-based segmentation model has excellent global recognition but poor thin-structure boundaries. Which design modification best addresses the mismatch?

57 A Vision Transformer uses patch size on an image of dimensions . Ignoring a class token, how does the token count change when the patch size is halved while and remain fixed?

58 Compared with a convolutional network, a vanilla ViT trained from scratch on a small image dataset often underperforms primarily because it has:

59 What is the key computational advantage of shifted-window attention in Swin Transformer over global self-attention for a feature map with tokens?

60 A Swin Transformer uses alternating regular and shifted windows. If shifted windows were removed while all other components remained unchanged, what capability would be most impaired?