search

syfafterzy / SVDNet-for-Pedestrian-Retrieval

Code for https://arxiv.org/abs/1703.05693
138 stars 47 forks source link

Image size in training #6

Open a462428 opened 6 years ago

a462428 commented 6 years ago

Hi,

Image size in Market1501 is 64*128

however, in SVD training we have to resize them to 227*227

I wanna know whether directly resizing data will improve the accuracy or not?

if not, what should I do?

ghost commented 6 years ago

First, the input size of SVDNet is following the backbone model, e.g., 224 \times 224 for ResNet-50 backbone and 227 \times 227 for CaffeNet backbone.

Second, the input size of model do impact the performance. Here have a paper (Beyond Part Models: Person Retrieval with Refined Part Pooling (and a Strong Convolutional Baseline)) gives experiment on the impact of input size, and you can see more details in this paper.

a462428 commented 6 years ago

Yes I know~

but I wanna check whether we can directly resize our image to 227*227 in training phase?

ghost commented 6 years ago

@a462428 Yes! caffe can resize the image during training.