【CV】What is the object detection model? Part3

Object detection explanation part3.

3. Evaluation metrics and loss functions

Evaluation metrics

The performance of object detection is measured by evaluation metrics such as IoU and mAP.
Object detection is a task to detect where something is.
IoU: How much the predicted region and correct region overlap
mAP: AP (average of average precision) for each class. AP is calculated from the area under the PR curve.

Loss function

Cross Entropy: Often used in classification
Focal Loss: When there is class imbalance or difficulty imbalance. Emphasizes loss of classes with a small number of classes or data that is easy to classify.
IoU: How much the predicted region and correct region overlap
GIoU: The size of the region that encompasses the predicted region and correct region. It becomes possible to take into account the degree of separation, such as off but close.

4. History of object detection models

Object detection technology has evolved from template matching and hand-crafted features to deep learning.
Template matching: Based on correlation with the template image
Hand-crafted features: Simple machine learning model using features such as SIFT and HOG. SIFT and HOG are put into SVM, etc. to classify background and objects.

Deep learning: NN

5. Base model

In recent years, high generalization performance has been achieved by using base models trained with large amounts of data. The base model can be used for various tasks without additional training. Fine tuning improves performance.

・Image system
Meta has released various base models for object detection.

SAM (2023):
High-precision segmentation is possible even with zero shots. Video support will be available in July 2024 (SAM2).

Depth Anything (2024):
Estimates depth. For example, whether a dog is in the foreground of the image or in the background.

Sapiens (2024):
The following four models are collectively called Sapiens. Trained with 300 million images.
• Sapiens Pose: 2D pose estimation
• Sapiens Segmentation: Segmentation of body parts
• Sapiens Depth: Depth estimation
• Sapiens Normal: Surface normal prediction