Determining the type of location where an image scene is located from an image is a common task in image understanding. In essence, when the scene category annotation data is sufficient, it can belong to a kind of image classification, so directly using the existing mature network architecture such as ResNet can realize the recognition of high-precision image-related places.
The practice of this article is taken from: http://places2.csail.mit.edu/download.html
The dataset covers 365 image scenes, and also provides pre-trained models of various network architectures, mainly as follows:
Pre-trained CNN models on Places365-Standard:
- AlexNet-places365: deploy weights
- GoogLeNet-places365: deploy weights
- VGG16-places365: deploy weights
- VGG16-hybrid1365: deploy weights
- ResNet152-places365 fine-tuned from ResNet152-ImageNet: deploy weights
- ResNet152-hybrid1365: deploy weights
- ResNet152-places365 trained from scratch using Torch: torch model converted caffemodel:deploy weights. It is the original ResNet with 152 layers. On the validation set, the top1 error is 45.26% and the top5 error is 15.02%.
- ResNet50-places365 trained from scratch using Torch: torch model. It is Preact ResNet with 50 layers. The top1 error is 44.82% and the top5 error is 14.71%.
- To use the alexnet and vgg16 caffemodels in Torch, use the torch library loadcaffe, where you could simply load the caffe model use the following commands. But note that the input image scale should be from 0-255, which is different to the 0-1 scale in the previous resnet Torch models trained from scratch in fb.resnet.torch.
2. Experimental results
Categorize the locations on the map as: bars, restaurants, or coffee shops.
This is a test photo in the dataset, defined as a meeting room.
The identification of this waiting room is also very accurate.
See: https://timgsa.baidu.com/timg?image&quality=80&size=b9999_10000&sec=1523878667027&di=287398ec5e55869341ba2747794612a3&imgtype=0&src=http%3A%2F%2Fimg.pconline.com.cn%2Fimages%2F8%2F0photoblog% 2F0%2F8700542%2F20094%2F30%2F1241086150942.jpg
Basketball courts are also in the top three
The port terminal is also in the top few.