计算机视觉新手疑问：为何CNN输入图像尺寸多为特定数值？

Great question—this is something tons of folks stumble on when they’re starting out with computer vision and deep learning. Let’s break down the reasoning behind these standard sizes, plus what happens when you deviate from them.

Key Reasons for Standard Input Sizes #

• Historical Dataset Precedents: Many of these sizes come from iconic datasets that shaped the field. For example, MNIST (the "hello world" of image classification) uses 28×28 grayscale images—so every beginner learns with this size, and it’s stuck around for quick prototyping. Similarly, early versions of ImageNet used 256×256 images, which became a baseline for models like AlexNet and VGG.
• Model Architecture Constraints: Traditional CNNs (especially older ones) rely on fixed-size layers, particularly fully connected layers at the end. For example, a VGGNet trained on 224×224 images uses 5 rounds of 2×2 pooling—224 divided by 2^5 equals 7, a nice integer that feeds directly into the fully connected layers. If you pick a size that doesn’t divide cleanly after pooling/convolution steps, you’ll end up with non-integer feature map dimensions, which either breaks the model or requires messy padding/downsampling workarounds.
• Computational Efficiency: Sizes that are powers of 2 (like 256, 512) play nicely with GPU hardware. GPUs are optimized to handle tensors with dimensions that are multiples of 2, as it simplifies memory allocation and parallel computation. Even 28×28, while not a power of 2, is small enough that the overhead is negligible for its use case (simple digit recognition).
• Pretrained Model Compatibility: Most state-of-the-art models are trained on standard sizes (e.g., 224×224 for ResNet, 512×512 for some segmentation models). Using the same input size lets you directly load pretrained weights and fine-tune them, which saves massive amounts of training time and data. Straying from these sizes means you’ll have to adjust the model (like replacing fixed fully connected layers with global average pooling) or retrain from scratch, which is rarely practical.

What Happens If You Use Arbitrary Image Sizes? #

The consequences depend on your model architecture and use case, but here are the most common issues:

• Architecture Breakage (Older Models): If your CNN has fixed fully connected layers, changing the input size will alter the dimensions of the final feature map. For example, if you swap a 28×28 input for 30×30 in a MNIST model, the final feature map after pooling might be 7.5×7.5 (instead of 7×7)—which can’t feed into a fully connected layer expecting a fixed number of neurons. Modern models often use global average pooling instead, which avoids this issue, but older architectures will throw errors.
• Performance Degradation: Even if the model runs, deviating too far from the pretrained size can hurt accuracy. Pretrained models learn features optimized for their original input size—using a much smaller size might strip away critical details, while a much larger size could introduce redundant noise the model wasn’t trained to handle. For tasks like object detection, anchor boxes (predefined bounding box sizes) are tuned for specific input dimensions, so arbitrary sizes can throw off box alignment and reduce detection precision.
• Computational Resource Shifts: Larger input sizes drastically increase memory usage and compute time. A 512×512 image has 4x the pixels of a 256×256 image, which means 4x more operations per batch—this can lead to GPU out-of-memory errors if you’re not careful. Smaller sizes are faster but may lack the detail needed for complex tasks like high-resolution segmentation.
• Edge Handling Anomalies: Odd-sized inputs can lead to inconsistent padding during convolution. For example, if you use a 3×3 kernel on a 29×29 image without padding, the output will be 27×27—this can cut off more edge information than expected, which might matter for tasks where edge details are critical (like medical imaging).

At the end of the day, you can use arbitrary sizes if your model is designed to handle them (e.g., using adaptive pooling, fully convolutional networks), but sticking to standard sizes will save you headaches, especially when working with pretrained models.

内容的提问来源于stack exchange，提问作者Yixuan Sun