Deep learning computer vision covers many tasks that help machines understand images and videos. Common tasks include image classification (deciding what the main object is), object detection (finding objects and drawing boxes around them), and segmentation (labeling each pixel or separating individual objects). It also includes keypoint detection for landmarks such as body joints, optical flow for estimating motion between frames, and video classification for understanding actions in videos.
In simple terms, these tasks move from answering “what is in the image?” to “where is it?”, “what shape is it?”, and “how is it moving?” .
Image classification predicts one or more labels for the whole image.
Examples:
- cat vs dog
- healthy leaf vs diseased leaf
- normal product vs defective product
This task is useful when global content matters more than object location.
Object detection predicts both category and location.
Examples:
- person detection
- helmet detection
- vehicle detection
Detection is very important because it becomes the basis for many downstream tasks such as counting, tracking, and event reasoning.
Segmentation works at pixel level.
Two common forms:
- semantic segmentation: classify each pixel by category
- instance segmentation: separate individual objects of the same category
Examples:
- road area segmentation
- product boundary extraction
- background removal
Tracking connects detections over time.
It answers a different question from detection:
- not only what is here now
- but which object remains the same across frames
This is essential for:
- counting people
- monitoring movement
- behavior analysis
Pose estimation predicts structured keypoints such as joints or landmarks.
Examples:
- human skeleton estimation
- hand landmarks
- animal pose estimation
This is useful when shape and posture matter more than a simple box.
Face recognition goes beyond detecting a face. It tries to identify or verify who the person is.
That usually involves:
- face detection
- face alignment
- feature extraction
- comparison or matching
This task introduces important concerns around privacy, bias, and security.
OCR extracts text from images or video.
Examples:
- receipts
- labels
- license plates
- forms
OCR is one of the clearest examples of vision interacting with language.
Some practical systems care less about static objects and more about events:
- line crossing
- intrusion
- fall detection
- loitering
These tasks often combine detection, tracking, and rule-based logic.
- "Detection is always enough."
- Some applications need segmentation, pose estimation, OCR, or temporal reasoning.
- "Tracking is just repeated detection."
- Tracking also requires identity continuity across time.
- "Face recognition and face detection are the same."
- Detection finds the face. Recognition identifies or verifies the person.
- Choose one application from each of the following areas and identify the most suitable vision task:
- smart retail
- factory inspection
- traffic analysis
- document processing
- Explain when segmentation is better than detection.
- Explain why tracking is important in video analytics but not necessary in every image task.
Deep learning computer vision includes many task types, each with its own outputs, strengths, and applications. Understanding this task map helps the learner choose the right tool for a problem. In the next section, the course uses YOLO26 as the main practical case for learning how to train and deploy a custom model.
Continue to 4.6 Train and Deploy Your Own Vision Model.
https://github.com/NVIDIA/semantic-segmentation
https://github.com/yehengchen/Object-Detection-and-Tracking
https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation