Saral Shiksha Yojna
Courses/Computer Vision

Computer Vision

CSE471
Prof. Makarand Tapaswi + Prof. Charu Sharma•Spring 2025-26•4 credits
Sample Papers/Mock Paper 15 — Edge Cases, Corner Cases, 'What happens if...'

Mock Paper 15 — Edge Cases, Corner Cases, 'What happens if...'

Duration: 150 min • Max marks: 100

Section A — Quick Edge Cases (2 marks each, 40 marks)

40 marks
  1. 1.What happens to a CNN's output if you pass a completely black (all zeros) image?2 m
  2. 2.IoU of two identical bounding boxes?2 m
  3. 3.IoU of two non-overlapping bounding boxes?2 m
  4. 4.What happens to BatchNorm at batch size = 1?2 m
  5. 5.What happens to Dropout at inference?2 m
  6. 6.Gradient of identity y = x?2 m
  7. 7.What happens to softmax with temperature τ → 0?2 m
  8. 8.What happens to softmax with τ → ∞?2 m
  9. 9.Attention output if all attention scores are equal?2 m
  10. 10.What happens to a U-Net if you remove all skip connections?2 m
  11. 11.CNN trained on MNIST, tested on 90° rotated MNIST — what happens?2 m
  12. 12.What happens to attention complexity if you double sequence length N?2 m
  13. 13.What if you apply NMS with threshold = 0?2 m
  14. 14.What if you apply NMS with threshold = 1?2 m
  15. 15.What happens to PCA when all data points are at the origin (zero variance)?2 m
  16. 16.Pose heatmap with TWO equal peaks — what does it mean?2 m
  17. 17.What happens in contrastive SSL if augmentations are too WEAK?2 m
  18. 18.What happens if DINO's output dimension is too small (e.g., 100)?2 m
  19. 19.What happens to MAE training if you use 15% mask (BERT-style) instead of 75%?2 m
  20. 20.What happens in 3DGS if adaptive density control is disabled?2 m

Section B — Detailed Edge Cases (5 marks each, 30 marks)

30 marks
  1. 1.What happens if you train CLIP with only 1000 image-text pairs?5 m
  2. 2.What happens if all Faster R-CNN anchors have the same aspect ratio?5 m
  3. 3.What happens in DINO if you use only ONE crop instead of multi-crop?5 m
  4. 4.What happens to ViT performance if you remove positional encodings?5 m
  5. 5.What happens to a GAN Generator if the Discriminator is 'perfect' (D=1 for real, D=0 for fake)?5 m
  6. 6.Your binary classifier achieves train 99.99%, val 65%, test 50% (chance). Describe what's happening and how to fix.5 m

Section C — Long-Form Corner Cases (10 marks each, 30 marks)

30 marks
  1. 1.Detect license plates that are only 5-10 pixels wide in a 4K image. Engineering challenges and complete solution.10 m
  2. 2.Train 99.99%, val 65%, test 50% — diagnose and fix (long form).10 m
  3. 3.Your CLIP gives suspiciously high 98% zero-shot accuracy on a classification task. Investigate whether this is too good to be true.10 m

Track your attempt locally — score and time are recorded in your browser. (Coming soon: timed-attempt mode.)