Computer Vision Limitations
Computer vision systems have real limitations and failure modes. Understanding these is essential for building responsible AI systems.
8 min•By Priygop Team•Updated 2026
Known Limitations
- Adversarial attacks: adding tiny invisible noise to an image can cause a classifier to completely misidentify it with high confidence
- Distribution shift: a model trained on studio photos of products may fail on blurry phone photos
- Bias: models trained mostly on photos of one demographic perform worse on others. This has been documented in face recognition
- Lack of common sense: a model can confidently classify a random pattern as a specific object because it matches learned visual patterns
- Occlusion: objects that are partially hidden can confuse detectors
- Poor lighting and weather: models trained on clear daylight photos may fail in fog, rain, or low light
Diagram
Loading diagram…
Deep Learning ⊂ Machine Learning ⊂ Artificial Intelligence
Key Takeaways from This Module
- Computer vision teaches AI to understand images, video, and camera feeds
- Images are number grids: grayscale (1 channel) or RGB (3 channels)
- Preprocessing: resize, normalize, and augment before feeding images to a model
- CNNs use sliding filters to detect local patterns at multiple scales
- Image classification: label the whole image. Object detection: find and label each object with a box
- Face recognition uses face embeddings and similarity comparison
- Real applications: medical imaging, self-driving, quality control, security, retail
- Limitations: adversarial attacks, bias, distribution shift, occlusion
Key Takeaways
- Computer vision systems have real limitations and failure modes.
- Adversarial attacks: adding tiny invisible noise to an image can cause a classifier to completely misidentify it with high confidence
- Distribution shift: a model trained on studio photos of products may fail on blurry phone photos
- Bias: models trained mostly on photos of one demographic perform worse on others. This has been documented in face recognition