Custom detection dataset in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Custom detection dataset

Which metric matters for Custom Detection Dataset and WHY

For object detection tasks using a custom dataset, the key metrics are Precision, Recall, and Mean Average Precision (mAP). Precision tells us how many detected objects are actually correct, while Recall tells us how many real objects were found. mAP summarizes how well the model detects objects across different classes and confidence levels. These metrics matter because detection involves both locating and classifying objects, so we need to balance finding all objects (Recall) and avoiding false alarms (Precision).

Confusion Matrix for Object Detection

In detection, confusion matrix is more complex because predictions include bounding boxes and classes. But for a single class, it can be simplified as:

      +----------------+----------------+
      |                | Predicted Obj  |
      |                | Yes      No    |
      +----------------+----------------+
      | Actual Obj Yes | TP       FN    |
      | Actual Obj No  | FP       TN    |
      +----------------+----------------+
    

Where:

• TP: Correctly detected objects with good overlap (IoU > threshold)
• FP: Wrong detections or duplicates
• FN: Missed objects
• TN: Not usually counted in detection

Precision vs Recall Tradeoff with Examples

Imagine a security camera detecting people:

• High Precision, Low Recall: The camera only alerts when very sure. Few false alarms, but might miss some people.
• High Recall, Low Precision: The camera alerts on many things, catching almost all people but also many false alarms.

Choosing the right balance depends on the use case. For safety, high recall is better to not miss anyone. For convenience, high precision avoids false alerts.

Good vs Bad Metric Values for Custom Detection Dataset

• Good: Precision and Recall above 0.8, mAP above 0.75 means the model detects most objects correctly and rarely makes mistakes.
• Bad: Precision or Recall below 0.5 means many false detections or many missed objects. mAP below 0.4 shows poor overall detection quality.

Common Pitfalls in Detection Metrics

• Accuracy Paradox: High accuracy can be misleading if most images have no objects. The model can guess "no object" and be right often but useless.
• Data Leakage: Using test images in training inflates metrics falsely.
• Overfitting: Very high training metrics but low test metrics means the model memorizes training data and won't generalize.
• Ignoring IoU Threshold: Counting detections with low overlap as correct inflates metrics unfairly.

Self Check

Your detection model has 98% accuracy but only 12% recall on detecting cars. Is it good for production?

Answer: No. The high accuracy is misleading because most images might have no cars, so guessing "no car" is often right. The very low recall means the model misses most cars, which is bad for detection. You should improve recall before using it in production.

Key Result

Precision, Recall, and mAP are key to evaluate detection models; balance between them shows true performance.

Practice

1. What is the main purpose of creating a Custom detection dataset in PyTorch?

easy

A. To perform image classification without labels

B. To convert images to grayscale only

C. To generate random images for testing

D. To load images and their bounding boxes with labels for training detection models

Solution

1. Step 1: Understand dataset role in detection

   A detection dataset provides images along with bounding boxes and labels for each object to train detection models.

2. Step 2: Identify correct purpose

   Classification without labels, random image generation, and grayscale conversion do not provide bounding boxes or labels needed for detection tasks.

3. Final Answer:

   To load images and their bounding boxes with labels for training detection models -> Option D

4. Quick Check:

   Detection dataset = images + bounding boxes + labels [OK]

Hint: Detection datasets always include bounding boxes and labels [OK]

Common Mistakes:

• Thinking detection datasets only have images
• Confusing detection with classification datasets
• Ignoring bounding boxes in dataset

2. Which of the following is the correct way to define the __getitem__ method in a custom detection dataset class?

easy

A. def __getitem__(self): return image, label

B. def __getitem__(self, idx): return image, label

C. def __getitem__(self, idx): return image, bounding_boxes, labels

D. def getitem(self, idx): return image, bounding_boxes

Solution

1. Step 1: Recall __getitem__ signature

   The method must accept an index parameter and return image, bounding boxes, and labels for detection.

2. Step 2: Check options for correct syntax and return values

   def __getitem__(self, idx): return image, bounding_boxes, labels correctly uses def __getitem__(self, idx): and returns image, bounding boxes, and labels. Others miss parameters or return incomplete data.

3. Final Answer:

   def __getitem__(self, idx): return image, bounding_boxes, labels -> Option C

4. Quick Check:

   __getitem__ needs idx and returns image + boxes + labels [OK]

Hint: __getitem__ always takes index and returns image with boxes and labels [OK]

Common Mistakes:

• Omitting the idx parameter
• Returning only image and label without boxes
• Using wrong method name without underscores

3. Given this snippet from a custom detection dataset, what will be the output type of dataset[0]?

class CustomDataset(torch.utils.data.Dataset):
    def __getitem__(self, idx):
        image = self.images[idx]
        boxes = self.boxes[idx]
        labels = self.labels[idx]
        target = {"boxes": boxes, "labels": labels}
        return image, target

medium

A. A tuple containing an image tensor and a dictionary with boxes and labels

B. A list containing image, boxes, and labels separately

C. Only the image tensor

D. A dictionary with keys 'image' and 'target'

Solution

1. Step 1: Analyze return statement

   The method returns image, target where target is a dictionary with keys 'boxes' and 'labels'.

2. Step 2: Determine output type

   Returning two items separated by comma means a tuple is returned: (image, target dictionary).

3. Final Answer:

   A tuple containing an image tensor and a dictionary with boxes and labels -> Option A

4. Quick Check:

   Return image, target dict = tuple [OK]

Hint: Comma-separated returns form a tuple in Python [OK]

Common Mistakes:

• Thinking return is a list instead of tuple
• Assuming only image is returned
• Confusing dictionary keys with return structure

4. You wrote this __getitem__ method but get an error when loading data:

def __getitem__(self, idx):
    image = self.images[idx]
    boxes = self.boxes[idx]
    labels = self.labels[idx]
    target = {"boxes": boxes, "labels": labels}
    return image, target

What is the most likely cause of the error?

medium

A. Missing return statement

B. boxes and labels are not converted to tensors

C. image is not normalized

D. Incorrect method name without underscores

Solution

1. Step 1: Check common dataset errors

   Detection models expect bounding boxes and labels as tensors, not raw lists or arrays.

2. Step 2: Identify missing tensor conversion

   If boxes and labels are not converted to tensors, PyTorch will raise errors during training or data loading.

3. Final Answer:

   boxes and labels are not converted to tensors -> Option B

4. Quick Check:

   Boxes and labels must be tensors [OK]

Hint: Always convert boxes and labels to tensors before returning [OK]

Common Mistakes:

• Forgetting to convert boxes and labels to torch tensors
• Assuming image normalization fixes tensor errors
• Using wrong method name causing attribute errors

5. You want to add a transform that flips images horizontally and adjusts bounding boxes accordingly in your custom detection dataset. Which approach is best?

hard

A. Apply the horizontal flip transform to both image and bounding boxes inside __getitem__

B. Flip bounding boxes but not the image

C. Apply the transform only after training the model

D. Apply the horizontal flip transform only to the image inside __getitem__, leave boxes unchanged

Solution

1. Step 1: Understand transform requirements

   When flipping images, bounding boxes must be adjusted to match the new image layout.

2. Step 2: Apply transform to both image and boxes

   Applying the transform only to the image or only to boxes causes mismatch and incorrect training data.

3. Final Answer:

   Apply the horizontal flip transform to both image and bounding boxes inside __getitem__ -> Option A

4. Quick Check:

   Flip image and boxes together for correct data [OK]

Hint: Transform image and boxes together to keep alignment [OK]

Common Mistakes:

• Flipping only image or only boxes
• Applying transforms after training
• Ignoring bounding box coordinate updates