Source-linked AI summary
Places205-VGGNet Models for Scene Recognition
Limin Wang, Sheng Guo, Weilin Huang, Yu Qiao
TL;DR
Directly adapting ImageNet-trained VGGNet models does not yield good scene-recognition performance. This report trains VGGNet models on Places205 using Multi-GPU Caffe and finds state-of-the-art performance across three scene-recognition benchmarks.
Problem
Directly adapting VGGNet models trained on ImageNet does not yield good performance for scene recognition, while training deeper VGGNets on large-scale scene data is computationally demanding.
Method
The report trains VGGNet-11, VGGNet-13, and VGGNet-16 on Places205 using a Multi-GPU extension of Caffe, then transfers learned representations to other scene datasets.
Results
The trained Places205-VGGNet models achieve state-of-the-art performance on Places205, SUN397, and MIT67.
Takeaways & Limitations
The released Places205-VGGNet models are intended to advance further research in scene recognition.
Abstract
from arXiv · showhide
VGGNets have turned out to be effective for object recognition in still images. However, it is unable to yield good performance by directly adapting the VGGNet models trained on the ImageNet dataset for scene recognition. This report describes our implementation of training the VGGNets on the large-scale Places205 dataset. Specifically, we train three VGGNet models, namely VGGNet-11, VGGNet-13, and VGGNet-16, by using a Multi-GPU extension of Caffe toolbox with high computational efficiency. We verify the performance of trained Places205-VGGNet models on three datasets: MIT67, SUN397, and Places205. Our trained models achieve the state-of-the-art performance on these datasets and are made public available.
1. Introduction
The report addresses the poor performance of ImageNet-trained VGGNets when directly adapted to scene recognition by training VGGNet models on Places205. It releases the models and transfers their representations to other scene datasets.
- Motivation: ImageNet-trained VGGNets cannot be directly adapted to scene recognition with good performance.Training deeper VGGNets on a large-scale scene dataset is also computationally demanding and requires substantial training expertise.
- Approach: The report trains high-performance VGGNet models for scene recognition on the Places205 dataset.
- Contributions: The models are transferred to MIT67 and SUN397, where they achieve the best performance among the compared models.The report also extracts ConvNet features off-the-shelf for transfer.
2. Implementation Details
The authors train 11-, 13-, and 16-layer VGGNets with Caffe and multi-GPU computation, using staged initialization, augmentation, regularization, and multi-view testing. The implementation addresses the models’ high computational and memory demands.
- Network training: The implementation trains 11-, 13-, and 16-layer VGGNets, increasing depth while using the pretrained 11-layer model to initialize deeper networks.The original VGGNet design uses small 3 × 3 convolutional kernels, 1 × 1 strides, and 2 × 2 pooling windows.
- Computational setup: Multi-GPU Caffe is used because VGGNets have substantially higher computational cost and memory consumption than architectures such as GoogLeNet.The extension also supplies corner-cropping and multi-scale-cropping augmentation techniques, which the implementation adopts.
- Optimization: Training uses mini-batch gradient descent with batch size 256, momentum 0.9, weight decay 0.0005, and dropout ratio 0.5.Images are resized to 256 × 256, while multi-scale crops are selected from widths and heights {256, 224, 198, 168} and resized to 224×224.
- Network training: The deeper VGGNet models initialize their first four convolutional layers and first two fully connected layers from the pretrained 11-layer model.
- Evaluation protocol: Testing averages predictions from 10 views formed by five 224 × 224 crops and their horizontal flips.The five crops are taken from the four corners and center of a 256 × 256 image.
3. Experiments
The experiments evaluate Places205-VGGNet models on Places205, MIT67, and SUN397. VGGNets outperform the compared architectures on Places205, while transferred representations achieve the best performance on MIT67 and SUN397.
- Experimental setup: The experiments evaluate the trained models on Places205, MIT67, and SUN397 using four GTX Titan-X GPUs for training.Training VGGNet-16 takes around 2 weeks.
- Places205: On Places205, VGGNets obtain much better performance than AlexNet, GoogLeNet, and CNDS-8.The results are summarized in Table 1.
- Transfer evaluation: For MIT67 and SUN397, the learned representations are transferred by extracting normalized fc6 features and training linear SVM classifiers.
- Transfer evaluation: The Places205-VGGNet models achieve the best performance on MIT67 and SUN397 among the compared public models.The comparison is reported in Table 2.
4. Conclusions
The report trains VGGNets on Places205 with a multi-GPU Caffe extension and reports state-of-the-art performance across three scene-recognition benchmarks. The trained models are released for further research.
- Conclusion: The Places205-VGGNet models achieve state-of-the-art performance on Places205, SUN397, and MIT67.
- Conclusion: The trained Places205-VGGNet models are released for further scene-recognition research.