Note_ Can Semantic Labeling Methods Generalize to Any City The Inria Aerial Image Labeling Benchmark

基本信息

2017 IGARSS （頂會）

Can Semantic Labeling Methods Generalize to Any City? The Inria Aerial Image Labeling Benchmark

筆記

做者的認爲如今遙感領域的算法受限於數據集。

• 數據集所涵蓋的面積比較小，遙感數據和地點關係比較大，因此算法的泛化能力也受到了數據集的限制。

  those images cover limited geographic areas and the evaluation procedure does not assess how the methods generalize to different contexts or more abstract semantic classes.

  > the image tiles tend to be self-similar and with uniform color histograms

因此，提出一個開放的數據集合：

Dataset features:

• Coverage of 810 km² (405 km² for training and 405 km² for testing)
• Aerial orthorectified color imagery with a spatial resolution of 0.3 m
• Ground truth data for two semantic classes: building and not building (publicly disclosed only for the training subset)

具體以下：

同時開放一個檢測平臺contest，提供測試集的測試服務,也是一個比賽。

做者的另外一個貢獻是，本身作了實驗，定了一個baseline，

實驗

第一步，將訓練集合分紅訓練集合和驗證集合，也就是small vallidation set。

先是作了base-FCN的實驗，而後參考論文（Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat, and Pierre Alliez, 「High-resolution semantic labeling with convolutional neural networks,」 arXiv preprint arXiv:1611.01962, 2016. ）結合各層特徵，作了Skip 的實驗。本身再修正，重點介紹了關於MLP的實驗。

主要的改進是Concatenate各個特徵層，而後，利用一個只有一個hidden層MLP來，實現分類。

總結

整個測試，注重兩個指標：

1. First, the accuracy,defined as the percentage of correctly classified pixels.
2. Secondly, the intersection over union (IoU) of the positive (building) class.

關於IOU的提高空間還很大~

The MLP network reaches about 60% IoU on the entire test set. This means that the output objects overlap the real ones by 60%, as assessed over a significant amount of test data. While there is certainly room for improvement·····