Saral Shiksha Yojna
Courses/Computer Vision

Computer Vision

CSE471
Prof. Makarand Tapaswi + Prof. Charu Sharma•Spring 2025-26•4 credits

Answer Structure Templates

University exams reward formatting. Use these.

Derivation (variance / scaling)

  1. State assumptions about input distribution (zero mean, unit variance).
  2. Compute E[·] and Var[·] of the intermediate expression.
  3. Show how the scaling factor (e.g., √dₖ) cancels the variance growth.
  4. Conclude with the operating regime where softmax / gradients are stable.
  5. Verify with a small numerical example.

Algorithm trace (NMS / mAP)

  1. Sort detections by score descending.
  2. Initialize empty kept-list and write the first column.
  3. Walk row by row: keep or suppress; for mAP, mark TP/FP and update cumulative P, R.
  4. Tabulate intermediate values — graders give partial credit for the table.
  5. State the final answer with units.

Architecture compare (e.g., RoI Pool vs RoI Align)

  1. One-line definition of each.
  2. Side-by-side comparison table (≥ 5 rows: quantization, gradient quality, mask AP, output shape, when to use).
  3. State the failure mode of the inferior option with a concrete example.
  4. Conclude with which to use and why.

Mechanism + 'why does it work?'

  1. State the mechanism formally (equation if applicable).
  2. Identify the failure mode of naïve alternatives.
  3. Connect the mechanism's term-by-term effect to that failure mode.
  4. Empirical evidence: cite a benchmark / paper result.
  5. One-sentence summary.

Case study / paper design

  1. Restate the input → output spec in your own words.
  2. Sketch the architecture as a 3-pillar diagram (encoder → connector → decoder/head).
  3. List the training signal and loss; if multi-task, give the weighted sum.
  4. Note evaluation metrics and dataset choice.
  5. Identify failure modes and at least one ablation you'd run.

Multi-part 'compare and contrast'

  1. Brief shared definition (1 sentence).
  2. Table of differences (≥ 6 rows).
  3. Quantitative comparison (params, FLOPs, accuracy on a standard benchmark).
  4. Which-to-use-when guidance.
  5. One-line conclusion.

Numerical (params / sequence length / output shape)

  1. List given values.
  2. Write the formula symbolically.
  3. Substitute numbers.
  4. Show intermediate arithmetic (per-layer × #layers).
  5. Box final answer with units.