CLIP (vision-language model) in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - CLIP (vision-language model)

Which metric matters for CLIP and WHY

CLIP matches images and text. The key metric is zero-shot accuracy. It shows how well CLIP picks the right text label for an image without extra training. This tells us if CLIP understands the connection between pictures and words.

Other useful metrics include Recall@K, which checks if the correct label is in the top K guesses. This matters because CLIP often suggests several possible matches.

Confusion matrix or equivalent visualization

CLIP's output is a similarity score between images and text. We can build a confusion matrix by treating the highest scoring text as the prediction for each image.

      | Predicted: Cat | Predicted: Dog | Predicted: Car |
      |----------------|----------------|----------------|
      | Actual: Cat    |      85        |       10       |       5        |
      | Actual: Dog    |      8         |       90       |       2        |
      | Actual: Car    |      3         |       7        |       90       |
    

This shows how often CLIP correctly matches images to their true labels (diagonal numbers) versus mistakes (off-diagonal).

Precision vs Recall tradeoff with examples

For CLIP, precision means when it says an image matches a text, how often is it right? Recall means how many true matches it finds out of all possible matches.

Example: If CLIP is used to find images of "dogs" in a big gallery, high recall means it finds most dog images. High precision means most images it finds are really dogs.

If you want to avoid showing wrong images (like cats labeled as dogs), prioritize precision. If you want to find all dog images even if some mistakes happen, prioritize recall.

What "good" vs "bad" metric values look like for CLIP

Good: Zero-shot accuracy above 70% on standard datasets means CLIP understands image-text links well. Recall@5 above 90% means the right label is almost always in the top 5 guesses.

Bad: Accuracy below 50% means CLIP struggles to match images and text. Low recall means it misses many correct matches, making it unreliable for search or classification.

Common pitfalls in CLIP metrics

• Accuracy paradox: High accuracy can happen if many images belong to one class, hiding poor performance on others.
• Data leakage: Using test images or captions seen during training inflates metrics falsely.
• Overfitting: Fine-tuning CLIP on small datasets can reduce generalization, hurting zero-shot ability.
• Ignoring top-K metrics: Only checking top-1 accuracy misses how well CLIP ranks relevant labels.

Self-check question

Your CLIP model has 98% accuracy but only 12% recall on a rare class like "fire trucks." Is it good for production? Why or why not?

Answer: No, it is not good. The high accuracy likely comes from many common classes, but the very low recall on "fire trucks" means CLIP misses most fire truck images. For rare but important classes, recall matters more to avoid missing them.

Key Result

Zero-shot accuracy and Recall@K are key metrics showing how well CLIP matches images to text without extra training.

Practice

1. What is the main purpose of the CLIP model in computer vision?

easy

A. To connect images and text by learning their relationship

B. To generate images from random noise

C. To classify images into fixed categories without text

D. To detect objects using bounding boxes only

Solution

1. Step 1: Understand CLIP's design goal

   CLIP is designed to learn how images and text relate to each other, enabling it to match images with descriptions.

2. Step 2: Compare options with CLIP's purpose

   Options A, B, and D describe other tasks like classification without text, image generation, or object detection, which are not CLIP's main function.

3. Final Answer:

   To connect images and text by learning their relationship -> Option A

4. Quick Check:

   CLIP links images and text = C [OK]

Hint: CLIP matches images with text descriptions [OK]

Common Mistakes:

• Confusing CLIP with image generation models
• Thinking CLIP only classifies images
• Assuming CLIP detects objects with bounding boxes

2. Which of the following is the correct way to load a pre-trained CLIP model using Python's transformers library?

easy

A. model = transformers.CLIP('openai/clip-vit-base-patch32')

B. model = CLIP.from_pretrained('clip-base')

C. model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32')

D. model = load_clip('vit-base')

Solution

1. Step 1: Recall the transformers library syntax

   The correct method to load a pre-trained model is using the class name with from_pretrained and the model identifier string.

2. Step 2: Match options to correct syntax

   model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32') uses CLIPModel.from_pretrained with the correct model name. Others use incorrect class names or methods.

3. Final Answer:

   model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32') -> Option C

4. Quick Check:

   Use CLIPModel.from_pretrained() with model name [OK]

Hint: Use CLIPModel.from_pretrained('model-name') to load CLIP [OK]

Common Mistakes:

• Using wrong class names like CLIP or transformers.CLIP
• Missing from_pretrained method
• Using incomplete or incorrect model identifiers

3. Given the following Python code snippet using CLIP, what will be the output type of image_features?

from transformers import CLIPProcessor, CLIPModel
from PIL import Image

model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32')
processor = CLIPProcessor.from_pretrained('openai/clip-vit-base-patch32')

image = Image.new('RGB', (224, 224), color='red')
inputs = processor(images=image, return_tensors='pt')

outputs = model.get_image_features(**inputs)
image_features = outputs.detach().numpy()

medium

A. A numpy array representing the image embedding vector

B. A PIL Image object

C. A PyTorch tensor with gradients enabled

D. A string describing the image

Solution

1. Step 1: Understand model.get_image_features output

   This method returns a PyTorch tensor representing the image embedding vector.

2. Step 2: Analyze the conversion to numpy array

   Calling detach().numpy() converts the tensor to a numpy array without gradients, so the final type is a numpy array.

3. Final Answer:

   A numpy array representing the image embedding vector -> Option A

4. Quick Check:

   Image features output = numpy array [OK]

Hint: detach().numpy() converts tensor to numpy array [OK]

Common Mistakes:

• Thinking output is still a tensor with gradients
• Confusing image features with image object
• Expecting a text description instead of embeddings

4. Identify the error in this CLIP usage code snippet and select the fix:

from transformers import CLIPProcessor, CLIPModel

model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32')
processor = CLIPProcessor.from_pretrained('openai/clip-vit-base-patch32')

inputs = processor(text='a photo of a cat')
outputs = model.get_text_features(inputs)

medium

A. Change processor(text='a photo of a cat') to processor(text=['a photo of a cat'])

B. Change model.get_text_features(inputs) to model.get_text_features(**inputs)

C. Add return_tensors='pt' inside the processor call

D. Replace CLIPModel with CLIPTextModel

Solution

1. Step 1: Check how model methods accept inputs

   CLIP model methods expect keyword arguments unpacked from the processor output, so **inputs is needed.

2. Step 2: Identify the error and fix

   Passing inputs directly causes an error; changing to model.get_text_features(**inputs) fixes it.

3. Final Answer:

   Change model.get_text_features(inputs) to model.get_text_features(**inputs) -> Option B

4. Quick Check:

   Use **inputs to unpack processor output [OK]

Hint: Unpack processor output with ** when calling model methods [OK]

Common Mistakes:

• Passing processor output without unpacking
• Not using return_tensors='pt' in processor
• Confusing CLIPModel with CLIPTextModel

5. You want to find the most relevant image from a list using CLIP given a text query. Which approach correctly combines image and text features to find the best match?

hard

A. Match images by comparing their file names with the text query

B. Compare raw pixel values of images with text token IDs directly

C. Use Euclidean distance between unnormalized image and text features without preprocessing

D. Compute cosine similarity between normalized image and text feature vectors, then select the highest score

Solution

1. Step 1: Understand CLIP feature comparison

   CLIP produces feature vectors for images and text; similarity is measured by cosine similarity after normalization.

2. Step 2: Evaluate options for matching

   Compute cosine similarity between normalized image and text feature vectors, then select the highest score correctly uses cosine similarity on normalized vectors. Options B, C, and D use invalid or irrelevant methods.

3. Final Answer:

   Compute cosine similarity between normalized image and text feature vectors, then select the highest score -> Option D

4. Quick Check:

   Use cosine similarity on normalized features [OK]

Hint: Normalize features and use cosine similarity to match [OK]

Common Mistakes:

• Comparing raw pixels with text tokens
• Using Euclidean distance without normalization
• Matching based on file names instead of features