Computer Vision Application Development
What this domain covers
CV questions cover classification, detection, and segmentation pipelines, data augmentation, transfer learning, and deployment on ModelArts, Atlas, and HiLens.
Core rules and concepts
- Classification predicts one label per image; detection predicts boxes plus classes (two-stage Faster R-CNN; one-stage YOLO/SSD); semantic segmentation labels every pixel, instance segmentation separates objects.
- Backbones: ResNet for accuracy, MobileNet for edge efficiency.
- Augmentation: random crop, flip, rotation, color jitter, mosaic. Apply it to training data only; validation and test use deterministic resize and normalization.
- NMS removes duplicate boxes: sort by score, keep the top box, suppress boxes with IoU above the threshold.
- mAP is mean average precision across classes at IoU thresholds such as 0.5.
- Huawei edge deployment: convert the model, optimize with CANN/ATC, and deploy to Atlas or HiLens.
Worked example
A convolution takes a 224×224 input with kernel 7, stride 2, padding 3. Output = floor((224 + 2×3 − 7)/2) + 1 = floor(223/2) + 1 = 111 + 1 = 112. Reasoning: padding adds 6 pixels to the spatial size, the kernel consumes 7, and stride 2 divides the remaining span; floor ensures the partial trailing window is dropped. The result is 112×112, matching the ResNet stem.
Traps
Augmenting the validation set; reversing the NMS IoU condition; reporting mAP at a different IoU than specified; forgetting that semantic segmentation needs a background class; feeding BGR data to a model trained on RGB.