Open tkuri opened 2 years ago
単眼Depth推定(SIDE)モデルのいくつかの隠れ層のUnitは特定のDepthの範囲にのみ反応すること(Selectivity)を発見。そこでSIDEの解釈可能性をそのDepthのSelectivityによって定量化することを提案。更に元のアーキテクチャを変更せずに解釈可能なモデルを学習させる事を提案。
Depth推定の精度を損なうことなく(向上させることなく)、SIDE用のネットワークの解釈可能性を効果的に向上させることを確認。また深度補完にも適用できることを確認。
https://openaccess.thecvf.com/content/ICCV2021/html/You_Towards_Interpretable_Deep_Networks_for_Monocular_Depth_Estimation_ICCV_2021_paper.html
https://github.com/youzunzhi/InterpretableMDE
論文概要
単眼Depth推定(SIDE)モデルのいくつかの隠れ層のUnitは特定のDepthの範囲にのみ反応すること(Selectivity)を発見。そこでSIDEの解釈可能性をそのDepthのSelectivityによって定量化することを提案。更に元のアーキテクチャを変更せずに解釈可能なモデルを学習させる事を提案。
Depth推定の精度を損なうことなく(向上させることなく)、SIDE用のネットワークの解釈可能性を効果的に向上させることを確認。また深度補完にも適用できることを確認。
https://openaccess.thecvf.com/content/ICCV2021/html/You_Towards_Interpretable_Deep_Networks_for_Monocular_Depth_Estimation_ICCV_2021_paper.html
Code
https://github.com/youzunzhi/InterpretableMDE