Let me walk you through this clearly, based on hands-on experience with the TensorFlow Object Detection API:

What does first_stage_features_stride:16 mean? #

This parameter is specific to two-stage detectors like Faster R-CNN (the architecture implied by your "first stage" reference). Here's the breakdown:

• It defines the downsampling stride between your input image and the feature map used by the Region Proposal Network (RPN)—the "first stage" of the detector.
• For your 4464x2976 input images, a stride of 16 means the RPN operates on a feature map sized 4464/16 = 279 pixels wide and 2976/16 = 186 pixels tall.
• Each pixel in this feature map corresponds to a 16x16 pixel region in your original input. The RPN generates anchor boxes based on these feature map pixels, so this stride directly impacts how well the detector can "see" small objects in your images.

Should you adjust this parameter after resizing your input images? #

The short answer: it depends on the size of the objects you're trying to detect. Here's how to decide:

• If your target objects are relatively large (e.g., bounding box dimensions of 64x64 pixels or larger in the original input), keeping stride:16 is totally fine. The 16x16 receptive field per feature map pixel will still cover enough of the object to generate accurate proposals.
• If you have small objects (e.g., bounding boxes smaller than 32x32 pixels in the original input), consider reducing the stride to 8. A smaller stride creates a feature map twice as large (each pixel maps to an 8x8 input region), giving the RPN a finer-grained view—critical for detecting tiny objects that might get lost in a 16x16 stride's receptive field.
• Key tradeoff to note: Lowering the stride to 8 will significantly increase computational load. Your training and inference speeds will slow down because the feature map becomes 4x larger (width and height both doubled), so the RPN has four times as many pixels to process. Only make this change if small object detection is a top priority.

A quick sanity check: Pull up your XML annotations and calculate the average width/height of your bounding boxes relative to your 4464x2976 input. If most objects are tiny compared to the image size, adjusting the stride is worth testing. Otherwise, stick with 16 to keep performance efficient.

内容的提问来源于stack exchange，提问作者Ajinkya