Recap and final exam

Ten modules to go from a raw crossroads frame to a real-time detector that counts vehicles and pedestrians and defends its own accuracy number. Here is the course condensed, and the threads running through it.

The course at a glance

Module	What to retain
1. Three families of tasks	Classification, detection, segmentation form a hierarchy of increasing information and cost; reverse-engineer the label from the actual downstream decision
2. Two-stage detection (R-CNN)	Split where from what; Faster R-CNN replaced hand-crafted proposals with a learned RPN; RoI Align matters most for segmentation
3. One-stage detection (YOLO, SSD)	Dense predictions on a grid in a single pass; modern medium YOLOs match Faster R-CNN accuracy at 5 to 10x the speed
4. Anchors, NMS and thresholds	Anchors are shape priors; NMS collapses overlaps; confidence threshold governs precision-recall; eval and deployment use different values
5. Detection metrics	IoU is the atomic notion; AP is the area under the class PR curve; mAP@[0.5:0.95] is the primary metric; always report the size split
6. Semantic segmentation	Encoder-decoder with skip connections (U-Net) or atrous convolutions (DeepLab); mean IoU per class, never pixel accuracy on imbalanced data
7. Instance segmentation	Mask R-CNN adds a per-class mask head to Faster R-CNN; RoI Align was invented for this task; SAM makes annotation an order of magnitude cheaper
8. Multi-object tracking	Predict / associate / update loop; Kalman + Hungarian on IoU; ByteTrack leads on public benchmarks; watch IDF1, not just mAP
9. Annotation and augmentation	Inter-annotator IoU caps model quality; Albumentations for label-aware augmentation; near-duplicate leakage inflates mAP by tens of points
10. Project: custom detection	Video-aware split; YOLOv8 at higher image size for small objects; error analysis by class and by size; ONNX export with baked NMS

The threads running through the course

Two decisions dominate every vision project, and neither is architectural. The first is which task family to target: classification, detection or segmentation. Modules 1 and 9 both revisit it, because the annotation cost of choosing wrong is measured in months, not milliseconds. The second is how the train / test split is made: near-duplicate leakage (module 9) is the single most common cause of a model that ships with excellent metrics and disappoints in production. The architecture that closes those two decisions well outperforms a fancier architecture built on shaky data.

COCO metrics are more than a benchmark; they are a communication protocol. Modules 2, 3, 5, 7 and 10 all end with the same evaluator. Once every stakeholder in the project — modelling, product, safety — reads mAP@[0.5:0.95] with the AP_S / AP_M / AP_L split the same way, discussions stop being about aesthetics and start being about numbers. Skipping the split and quoting only the headline is the technical equivalent of quoting an average without a standard deviation.

Thresholds live in two worlds. Module 4 introduced the distinction, modules 5 and 10 lived it: at evaluation time the confidence threshold stays near zero so mAP can trace the full precision-recall curve; at deployment time the threshold is tuned to the asymmetric cost of your errors. Confusing the two situations is the number-one cause of the complaint "the model was great on the benchmark, it is unusable in production". The two thresholds are not the same knob at the same value with a different context; they are the same knob at different values on purpose.

The bottleneck at deployment is rarely the model. Modules 8, 9 and 10 spent as much time on labelling, tracking identity switches and post-processing NMS as on the detector itself. In the crossroads pipeline, a 5 % identity-switch rate silently invents 25 extra vehicles per 500; a leaked near-duplicate inflates test mAP by 15 points; a mismatched NMS between eval and deployment reintroduces duplicate counts. Whenever "the model is done and it still does not work", the problem is almost always in one of these three places.

The final exam

The exam has 40 questions covering all ten modules: distinguishing task families, reading COCO annotations, contrasting Faster R-CNN and YOLOv8, setting anchors and thresholds correctly, computing IoU by hand, reading a COCO report line by line, choosing between U-Net and DeepLab, deciding when Mask R-CNN is worth its cost, assembling a tracker with a counting line, spotting near-duplicate leakage, and diagnosing errors by class and object size.

Several questions present situations to diagnose: a model whose reported mAP looks fine but whose class-specific AP is catastrophic, a counting pipeline whose numbers drift up because of identity switches, a tracker that resets ids on every batch in production, an evaluation whose numbers were inflated by a video split that cut within a clip. Judgement is what gets assessed, not memorised API signatures.

On success, your certificate of completion is issued immediately; its number can be verified by any third party on the platform.

Before you start

Go back over the table above and, for each row, ask yourself "what symptom would I see if I got this wrong?". If you can explain why a low NMS threshold undercounts, why mAP alone hides catastrophic per-class failures, why a horizontal-flip augmentation may or may not be safe on your dataset, and why a Kalman-filter tracker without ByteTrack over-counts in fog, you are ready. Good luck!

Final exam

Ready to validate this course?

40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.

Start the exam

You need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.