Single-shot Models
YOLO vs SSD
| Feature | YOLO | SSD |
|---|---|---|
| Speed | Up to 155 FPS | 22-46 FPS |
| Accuracy | Generally lower, struggles with small objects | Higher, better with varying object sizes |
| Architecture | Fully connected layers, grid-based predictions | Convolutional layers, anchor boxes |
| Best Use Case | Real-time applications | Applications needing accuracy and speed balance |
Development of YOLO
- YOLO v2 (YOLO 9000) adds Batch Normalization, and starts to predict the anchor boxes via -means clusters; Huge dataset with object categories
- YOLO v3 - An Incremental Improvement makes detections at three different scales in the network, using a new backbone network called Darknet-53 with 53 convolutional layers
Two-stage Models
Traditional R-CNNs
- R-CNN first use selective search to find possible bounding boxes and converts detection task into classification task
- SPP apply spatial pooling to enable the classification CNNs to accept multiple-size inputs
- Fast R-CNN boost the classification speed by first convert the whole image into a feature map and then apply selective search directly on these features
- Faster R-CNN further abandons the selective search to use a RPN to find possible bounding boxes, and propose the anchor boxes method
- Mask R-CNN add a small FCN head to Faster R-CNN to output a segmentation prediction, originally for instance segmentation, but also improves the accuracy for detection task
Novel CNN-based Architectures
- R-FCN divides the class feature into parts and evaluate the scores for the certain part of the certain class on different positions of the whole image. Thus this eliminates the need to pass each RoI into a detector network.
- FPN proposes an additional top-down path to further enhance the semantic features.
Tackle the imbalance of foreground objects and background objects
- OHEM proposes a read-only copy of the RoI network to first do a loss prediction on all RoIs and then select the regions with least accuracy to the RoI network that enables back-propagation.
- RetinaNet invents Focal Loss to down-weight the loss assigned to well-classified examples
Transformer-based Architectures
DETR considers the detection task as a set prediction problem, invents an end-to-end model based on ViT that eliminates the need for Non-maximum Suppression and human designed anchor boxes