Machine Learning Computer Vision 4 — Questions and Answers
Question 1: Which network architecture is specifically designed for real-time object detection with a single forward pass?
- Faster R-CNN
- Mask R-CNN
- YOLO (You Only Look Once) (Correct answer)
- R-FCN
Correct answer: YOLO (You Only Look Once)
YOLO performs detection in a single network pass by dividing the image into a grid and predicting boxes and classes simultaneously, enabling real-time performance.
Question 2: What is the function of the encoder in an encoder-decoder architecture like U-Net?
- Upsampling feature maps to full resolution
- Generating anchor boxes for detection
- Progressively downsampling input to capture high-level semantic features (Correct answer)
- Applying softmax to produce final class probabilities
Correct answer: Progressively downsampling input to capture high-level semantic features
The encoder uses convolutional and pooling layers to reduce spatial dimensions while increasing feature depth, capturing abstract semantic representations.
Question 3: What is optical flow used for in computer vision?
- Segmenting objects by color
- Estimating the motion of objects between consecutive video frames (Correct answer)
- Detecting text in natural images
- Reconstructing 3D point clouds from stereo images
Correct answer: Estimating the motion of objects between consecutive video frames
Optical flow estimates apparent pixel-level motion between consecutive frames, capturing the velocity field of moving objects in video.
Question 4: Which technique allows a CNN trained on ImageNet to be adapted for a smaller domain-specific dataset with limited data?
- Knowledge distillation
- Transfer learning (Correct answer)
- Generative pre-training
- Ensemble learning
Correct answer: Transfer learning
Transfer learning reuses weights from a model pre-trained on a large dataset, fine-tuning them on the target task to achieve good performance with limited data.
Question 5: In the context of image recognition, what does 'top-5 accuracy' mean?
- The model's accuracy on the five hardest test images
- The fraction of samples where the true label appears in the model's five highest-probability predictions (Correct answer)
- Accuracy after five rounds of fine-tuning
- Average accuracy across five random train/test splits
Correct answer: The fraction of samples where the true label appears in the model's five highest-probability predictions
Top-5 accuracy counts a prediction as correct if the ground-truth class is among the five classes assigned the highest probabilities by the model.
Question 6: What is the primary advantage of depthwise separable convolutions used in MobileNet?
- They increase model accuracy on ImageNet benchmarks
- They significantly reduce computation and parameter count compared to standard convolutions (Correct answer)
- They eliminate the need for batch normalization
- They allow arbitrary input resolutions without resizing
Correct answer: They significantly reduce computation and parameter count compared to standard convolutions
Depthwise separable convolutions factorize a standard convolution into a depthwise and pointwise step, drastically reducing multiply-add operations and parameters.
Question 7: Which self-supervised learning technique for computer vision trains a model to maximize agreement between differently augmented views of the same image?
- Autoencoders
- Contrastive learning (e.g., SimCLR) (Correct answer)
- Generative adversarial training
- Knowledge distillation
Correct answer: Contrastive learning (e.g., SimCLR)
Contrastive learning methods like SimCLR pull representations of augmented views of the same image together while pushing apart views from different images.
Machine Learning Computer Vision 4
Which network architecture is specifically designed for real-time object detection with a single forward pass?