Which parameters use for multiclass object detectors?

Hello guys, I want to make 3 object detectors for 5 different classes, I have 1865 images for use and these are the detectors I want to train: YOLOv4, Faster R-CNN, SSD. Can someone tell me what network, what input size, what number of anchor boxes, which solver name in training options, how many epochs, what minibatchsize and others to use for best performance ? I will be so thankful if someone will give me some starting info what to use for each object detector… i need to make object detectors with high values of average precision … thanks for any info :)

Answer 1

[Diwakar Diwakar]

YOLOv4, Faster R-CNN, and SSD) based on their commonly used configurations:

1. YOLOv4:

• Network architecture: YOLOv4
• Input size: Typically, YOLOv4 uses an input size of 416x416 or 608x608 pixels. You can experiment with different sizes to find the optimal balance between speed and accuracy.
• Number of anchor boxes: YOLOv4 typically uses three anchor boxes per scale.
• Solver name: YOLOv4 uses the Darknet framework.
• Epochs: . You can start with a reasonable number, such as 100 epochs, and monitor the loss and performance metrics during training to determine if more epochs are necessary.
• Minibatch size: YOLOv4 often uses a minibatch size of 64 or 128. However, this can vary depending on your hardware resources and memory constraints.

2. Faster R-CNN:

• Network: You can use the TensorFlow or PyTorch framework.
• Input size: Faster R-CNN typically works well with input sizes of 600x600 or 800x800 pixels.
• Anchor boxes: Faster R-CNN utilizes predefined anchor boxes of different scales and aspect ratios. You can experiment with anchor configurations based on the characteristics of your dataset.
• Solver: For Faster R-CNN, you can use the stochastic gradient descent (SGD) optimizer with a learning rate of 0.001.
• Epochs: Start with around 100 epochs and observe the model's performance. Adjust the number of epochs as needed.
• Minibatch size: A common minibatch size for Faster R-CNN is 16.

3. SSD (Single Shot MultiBox Detector):

• Network: You can use the TensorFlow or PyTorch framework, which both have SSD implementations available.
• Input size: SSD works well with input sizes of 300x300 or 512x512 pixels.
• Anchor boxes: SSD employs default anchor boxes at various scales and aspect ratios. Common configurations include using multiple feature maps to generate anchors with different scales and aspect ratios.
• Solver: Similar to Faster R-CNN, you can use the stochastic gradient descent (SGD) optimizer with a learning rate of 0.001.
• Epochs: Start with around 100 epochs and adjust as needed based on the performance of the model.
• Minibatch size: A common minibatch size for SSD is 32.

These recommendations are general starting points, and you may need to fine-tune the hyperparameters based on your specific dataset and requirements

[Adrian Kleffler]

Thank you so much for response. As shown in examples on MathWorks page for yolov4 i used Csp-darknet53-coco and resnet50 for faster rcnn and ssd… is that okay? And on my laptop i can’t run training when i use more than 4 minibatchsize when using yolov4 and faster rcnn… is that a problem that i train these detectors with minibatchsize = 4? And for yolov4 i used solver ADAM and for faster rcnn and ssd i used solver SGDM… is that okay? Thanks for response