预训练模型后调整分辨率：YOLOv1模型转换的参数疑问

Great question! This is a super common point of confusion when adapting pre-trained classification models for object detection, so let’s break it down clearly using YOLOv1 as an example.

Why input resolution doesn’t change parameter count #

First, let’s get a key concept straight: convolutional layer parameters don’t depend on input image size at all. A convolutional kernel’s parameter count is determined only by its spatial size (e.g., 3x3), the number of input channels it’s reading from, and the number of output channels it produces.

For example, a 3x3 conv layer that takes 64 input channels and outputs 128 channels has 3*3*64*128 = 73,728 parameters. Feed it a 224x224 image, a 448x448 image, or even a 1024x1024 image—those parameters stay exactly the same. The only thing that changes is the size of the feature maps generated at each layer (448x448 input will make feature maps twice as big in each dimension compared to 224x224, since each stride-2 convolution halves the spatial size).

How YOLOv1 transitions from 224x224 pre-training to 448x448 detection #

Here’s the exact process outlined in the YOLOv1 paper:

• Reuse the pre-trained convolutional backbone: All the convolutional layers from the ImageNet classification model (trained on 224x224) are kept intact. These layers already learned general visual features—like edges, textures, and basic object parts—that are transferable to detection tasks, so we don’t want to waste that pre-trained knowledge.
• Swap out the classification head for a detection head: The original model ended with fully connected layers that output 1000 class probabilities (for ImageNet). For detection, we remove these and replace them with brand-new fully connected layers tailored to YOLO’s output format: 7x7x(2*5 + 20). This output represents 7x7 grid cells, each with 2 bounding boxes (each with x/y coordinates, width/height, and confidence) plus 20 class probabilities (for the VOC dataset). These new fully connected layers are randomly initialized—they start from scratch.
• Fine-tune the whole model on detection data with 448x448 inputs: Now we train the modified model on detection datasets (like VOC) using 448x448 images. The pre-trained convolutional layers adapt their features to the higher resolution and detection-specific patterns, while the new fully connected layers learn to map those features to accurate bounding boxes and class predictions.

Why scale up to 448x448? #

224x224 was the standard input size for ImageNet classification back then, but it’s too low-resolution for reliable object detection—small objects get lost or blurred at that size. By scaling up to 448x448, YOLOv1 can capture finer details and smaller targets, which is critical for good detection performance. And since we reuse the pre-trained convolutional layers, we don’t have to train the entire model from scratch, saving a ton of time and compute.

内容的提问来源于stack exchange，提问作者walkerlala