| Presentation | 2022-06-09 Visualization of decisions from CNN models trained on OpenStreetMap images labeled based on traffic accident data Kaito Arase, Zhijian Wu, Tsuyoshi Migita, Norikazu Takahashi, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | The authors have recently conducted training of Convolutional Neural Networks (CNNs) on OpenStreetMap images each of which is labeled as ``danger'' or ``safe'' based on traffic accident data. Although the trained CNNs determine whether each area is danger or safe from the map image correctly with a high probability, the reason for this and the basis for their decisions are not clear. In this report, we use a method called Grad-CAM to visualize the basis of CNNs' decisions after learning map images as described above. The visualization result of the Grad-CAM depends on the convolutional layer of the CNN it is applied. The closer the layer is to the output layer, the better the features can be captured, but the resolution of the visualization is lower. Conversely, the closer the layer is to the input layer, the higher the resolution of the visualization, but the less well the features are captured. By analyzing the visualization results of the Grad-CAM for different convolutional layers of different models, we show that there certainly exist convolutional layers suitable for visualization, and clarify some characteristics of the trained CNNs when making their decisions. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | deep learning / map image / OpenStreetMap / Grad-CAM / traffic accident data |
| Paper # | NLP2022-10,CCS2022-10 |
| Date of Issue | 2022-06-02 (NLP, CCS) |
| Conference Information | |
| Committee | CCS / NLP |
|---|---|
| Conference Date | 2022/6/9(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Megumi Akai(Hokkaido Univ.) / Akio Tsuneda(Kumamoto Univ.) |
| Vice Chair | Masaki Aida(TMU) / Hidehiro Nakano(Tokyo City Univ.) / Hiroyuki Torikai(Hosei Univ.) |
| Secretary | Masaki Aida(TDK) / Hidehiro Nakano(Shibaura Insti. of Tech.) / Hiroyuki Torikai(Sojo Univ.) |
| Assistant | Tomoyuki Sasaki(Shonan Instit. of Tech.) / Hiroyasu Ando(Tsukuba Univ.) / Miki Kobayashi(Rissho Univ.) / " Hiroyuki YASUDA(The Univ. of Tokyo) / Yuichi Yokoi(Nagasaki Univ.) / Yoshikazu Yamanaka(Utsunomiya Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Complex Communication Sciences / Technical Committee on Nonlinear Problems |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Visualization of decisions from CNN models trained on OpenStreetMap images labeled based on traffic accident data |
| Sub Title (in English) | |
| Keyword(1) | deep learning |
| Keyword(2) | map image |
| Keyword(3) | OpenStreetMap |
| Keyword(4) | Grad-CAM |
| Keyword(5) | traffic accident data |
| 1st Author's Name | Kaito Arase |
| 1st Author's Affiliation | Okayama University(Okayama Univ.) |
| 2nd Author's Name | Zhijian Wu |
| 2nd Author's Affiliation | Okayama University(Okayama Univ.) |
| 3rd Author's Name | Tsuyoshi Migita |
| 3rd Author's Affiliation | Okayama University(Okayama Univ.) |
| 4th Author's Name | Norikazu Takahashi |
| 4th Author's Affiliation | Okayama University(Okayama Univ.) |
| Date | 2022-06-09 |
| Paper # | NLP2022-10,CCS2022-10 |
| Volume (vol) | vol.122 |
| Number (no) | NLP-65,CCS-66 |
| Page | pp.pp.46-51(NLP), pp.46-51(CCS), |
| #Pages | 6 |
| Date of Issue | 2022-06-02 (NLP, CCS) |