Lecture 16 — Capstone Demos & Review
Lecture 16
Dr. Muhammad Kazim
Department of Intelligent Systems, University of Lahore
Demo-Day Logistics: Schedule & Format
- Each team/student gets a fixed slot (announced in advance): set-up → short talk → live demo → Q&A → peer review, timed strictly so every slot in the 1.5-hour lab session gets a fair share.
- Set-up happens before your slot starts — have the Pi 5 powered, network/peripherals connected, and the application already running in the background if boot time is long.
- If your slot is early, be ready before session start; if it is late, keep your demo running/stable in the meantime rather than restarting it.
What to Bring / Bring-Up Checklist
- The Raspberry Pi 5 with your project pre-installed, plus power supply, camera, and any accessories (AI HAT/Hailo, ESP32-CAM, monitor/HDMI or a laptop for remote display).
- A charged laptop with your codebase and slides, in case the Pi needs last-minute debugging.
- A backup screen recording of a successful run — live demos fail sometimes (network, lighting, hardware); a backup video lets you still show working results and keeps grading fair.
- Your written project report, printed or ready to open instantly.
Structuring Your Short Talk
- Problem — what task does your system solve, and why does it matter?
- Approach — which pipeline/model/tier did you choose, and why (tie back to the Appendix C tier-comparison thinking: ESP32 vs Pi CPU vs Pi+Hailo)?
- Architecture — capture → inference → tracking/logic → action/output, in one diagram or sentence.
- Live demo — run it on the actual Pi 5, on a live or realistic input.
- Results — report accuracy/mAP/IoU (whichever applies) and FPS/latency/power on-device, not just on a laptop.
- Limitations — what breaks it, and what you would fix with more time.
Keep the talk tight; most of your slot is demo + Q&A, not slides.
Lab Project Grading Rubric
Of the 30-mark Lab Project (capstone) component:
| Criterion | Weight |
|---|
| Technical difficulty & correctness | 30% |
| Runs on the Raspberry Pi 5 in real time (or a justified budget) | 25% |
| Experimental rigour — baselines, ablations, honest metrics | 20% |
| Report clarity & reproducibility | 15% |
| Presentation & demo | 10% |
“Runs in real time” means measured, on-device FPS/latency — not a number from a laptop or Colab.
Peer-Review Guidance: What to Ask
- Is the claimed accuracy/FPS measured on the Pi 5, on a held-out set, honestly reported (including failure cases)?
- Does the demo generalise beyond the exact clip/scene it was tuned on?
- What is the action layer — does a detection actually change what the system does (log, alert, actuate), or does the pipeline stop at "detect and display"?
- What would break this system in a different lighting condition, camera angle, or with more objects/people in frame?
- Give one specific, actionable piece of feedback — peer review only helps if it is concrete.
Common Demo-Day Failure Modes & Mitigations
| Failure | Mitigation |
|---|
| Wi-Fi/network drops mid-demo | Prefer a local display or a wired connection; have the backup recording ready |
| Lighting in the demo room differs from your training data | Test in the actual room beforehand; bring a portable light if your system is lighting-sensitive |
| Model loads slowly / long cold-start | Start the process before your slot; warm up the camera and first inference in advance |
| Hailo/accelerator not detected | Re-run hailortcli identify before your slot; have a CPU-only fallback path |
| Live camera has nothing interesting in frame | Bring a physical prop / phone video to guarantee a detectable target |
Final Exam Review: Weeks 1–4
Classical fundamentals — know the definitions, the pipeline, and the key formulas:
- Week 1: the intelligent-vision-system pipeline (acquire → process → analyse → decide → act); radiometry, lenses, aperture, depth of field; pinhole/thin-lens models; perspective projection; exposure & motion blur.
- Week 2: CCD vs CMOS; Bayer CFA & demosaicing; the ISP pipeline; rolling shutter; sampling & quantisation; colour spaces (RGB, HSV, YCbCr, Lab).
- Week 3: LUTs, gamma/log maps, histogram equalisation/CLAHE; global/Otsu/adaptive/Sauvola thresholding; image arithmetic & blending.
- Week 4: correlation vs convolution, separable kernels, integral images; sharpening/Laplacian; gradients (Sobel/Prewitt/Scharr); median/bilateral/guided filters.
Final Exam Review: Weeks 5–9
- Week 5: 1D/2D DFT, the convolution theorem, ideal/Butterworth/Gaussian filters; sampling theorem, aliasing, Gaussian/Laplacian pyramids.
- Week 6: noise models; inverse/Wiener filtering, PSNR/SSIM; erosion/dilation/opening/closing, connected components, skeletonisation.
- Week 7: Canny in depth, Hough transform, moments/Hu invariants; Harris/Shi–Tomasi corners, DoG blobs, FAST.
- Week 8: SIFT/SURF/ORB/BRIEF, ratio test, FLANN vs brute-force; homography by DLT, RANSAC/MSAC, panorama stitching.
- Week 9: region growing, k-means/mean-shift, SLIC, IoU/Dice; watershed, GrabCut, active contours; the bridge to learned segmentation.
Final Exam Review: Weeks 10–12
- Week 10: intrinsics/extrinsics, distortion models, Zhang's calibration, reprojection error; epipolar geometry, fundamental/essential matrices, SGBM disparity, depth from disparity.
- Week 11: brightness constancy, Lucas–Kanade, Horn–Schunck, pyramidal KLT; MOG2/KNN background subtraction, the Kalman filter, KCF/CSRT, SORT, MOT metrics.
- Week 12: sliding-window detection, HOG+SVM, Haar cascades, bag-of-visual-words; perceptron → MLP, back-propagation, SGD/momentum/Adam, regularisation & augmentation.
The exam is weighted toward weeks 10–16 — make sure these three weeks and the ones below are solid.
Final Exam Review: Weeks 13–15
- Week 13: convolution & pooling, receptive fields, batch norm; the LeNet→ResNet→MobileNet→EfficientNet lineage; two-stage vs one-stage detection; FCN/U-Net/DeepLab/Mask R-CNN; mAP; the 5-step paper-reading method.
- Week 14: accelerators (GPU/NPU/DSP/FPGA); TensorRT/TFLite/ONNX Runtime/OpenVINO/Hailo; PTQ vs QAT, INT8/FP16, pruning, distillation; OpenCV 5.0's new DNN engine, breaking changes, and migration strategy.
- Week 15: scaled dot-product & multi-head attention, positional encoding, the ViT patch/[CLS]/position-embedding pipeline; MAE and contrastive self-supervision; CLIP's contrastive loss and zero-shot use; event cameras; a SLAM overview; ethics/privacy/bias/safety.
Where to Go Next
- Follow-on courses/topics: robotics & SLAM, advanced deep learning, reinforcement learning, embedded systems / real-time systems.
- Keep reading papers using the 5-step method from Lecture 13 — it works on any architecture, not just U-Net and the Transformer.
- Contribute to open source: OpenCV, Ultralytics YOLO, Hugging Face — real-world codebases built by exactly this kind of applied CV work.
- Challenge tier: revisit the Advanced Project Bank's challenge-tier ideas (a custom visual-classification feature, video-based detection monitoring, a full end-to-end custom vision system) as a next personal project, beyond the graded capstone.
Closing
From photons hitting a sensor (Week 1) to a Vision Transformer explaining a scene (Week 15) to a system you built running live on your own Raspberry Pi 5 (today) — that is the whole arc of Intelligent Vision Systems.
Good luck on the Final Exam, and congratulations on the capstone.