The real graded Computer Vision tasks the portal assigns, grouped into three levels.
The portal assigns these Computer Vision tasks. Finish Level 1 before Level 2 and Level 2 before Level 3. Your mentor may mark some tasks as required for the certificate.
Level 1: Foundations (12 tasks)
Task
What to do
Images as numbers
In a repo named cv-internship, load an image with OpenCV and explain pixels, channels (BGR vs RGB), shape and data type with printed examples.
Basic image operations
Resize, crop, rotate, flip and convert an image to grayscale. Save each result and show them in one grid in a notebook.
Drawing and text
Draw rectangles, circles and text labels on an image, as an object detector would, and save the annotated image.
Histograms and contrast
Plot the histogram of a dark photo, then improve it with histogram equalization and CLAHE. Compare the results.
Filters and blur
Apply average, Gaussian, median and bilateral filters to a noisy image and explain which one removes salt-and-pepper noise best.
Edge detection
Use Sobel and Canny edge detection on 3 images and tune the Canny thresholds. Explain what the thresholds do.
Thresholding and contours
Count coins or objects in a photo using thresholding and contour detection, and draw a box around each one.
Webcam capture
Open your webcam with OpenCV, show the live frame rate on screen and save a frame when a key is pressed.
Colour detection
Track an object of one colour in a webcam video using HSV colour ranges and draw its path.
Face detection
Detect faces in photos and the webcam with OpenCV's Haar cascade or MediaPipe and count the faces found.
Build an image dataset
Collect or download 200 or more images for 3 classes from a free dataset (Kaggle or Roboflow Universe), split them into train, validation and test folders and note the licence.
Image classifier with a pretrained model
Classify 20 of your own photos with a pretrained ResNet or MobileNet in PyTorch or Keras and show the top-3 predictions for each.
Level 2: Building real features (10 tasks)
Task
What to do
Transfer learning
Fine-tune MobileNet or EfficientNet on your 3-class dataset. Report accuracy, a confusion matrix and 5 images it got wrong.
Data augmentation
Train the same model with and without augmentation (flip, rotate, colour jitter) and compare the validation accuracy curves.
Label images for detection
Label 150 or more images with bounding boxes in Label Studio, CVAT or Roboflow and export them in YOLO format.
Train an object detector
Train a YOLO model (Ultralytics) on your labelled data. Report mAP50, precision and recall and show detections on new images.
Real-time detection
Run your detector on webcam or video and show the frames per second. Try a smaller model size and compare speed and accuracy.
Object tracking and counting
Track people or vehicles across video frames (ByteTrack or DeepSORT) and count how many cross a line.
OCR
Read text from 10 photos of bills or name boards with Tesseract or EasyOCR. Pre-process the images and report how accuracy changed.
Image segmentation
Run a segmentation model (YOLO-seg, SAM or U-Net) to separate objects from the background and save the masks.
Pose estimation
Use MediaPipe Pose or YOLO-pose to count squats or push-ups from a video using joint angles.
Image search with embeddings
Use CLIP to embed 200 images and search them by text, like 'red car' or 'a classroom', and show the top results.
Level 3: Professional work and capstone (8 tasks)
Task
What to do
Deploy a vision API
Serve your detector as a FastAPI endpoint that takes an image and returns boxes as JSON. Add file size and type checks.
Web app for your model
Build a Streamlit or Gradio app where users upload a photo and see predictions, and host it free on Hugging Face Spaces.
Optimise for edge devices
Export your model to ONNX or TensorFlow Lite, quantize it, and compare size and speed with the original.
Error analysis
Find 20 failure cases of your model, group them by cause (lighting, size, blur, angle) and fix the biggest group with more data or augmentation.
Privacy and bias check
Write how your app handles faces and personal images (consent, storage, blurring) and test whether accuracy differs across skin tones or lighting.
Industrial use case
Build a defect-detection demo for a real local need (bottle caps, PCB solder, fabric flaws or fruit quality) with a dataset, model and results.
Document the model
Write a model card: data, training settings, metrics, known limits and intended use.
Capstone: vision product
Ship a complete computer vision app (attendance by face with consent, parking counter, crop disease checker or similar) with a live demo link, repo and a 2-minute video.
Internship