tkuri / papers

Summarizing the papers I have read (Japanese)
41 stars 1 forks source link

Towards Interpretable Deep Networks for Monocular Depth Estimation #441

Open tkuri opened 2 years ago

tkuri commented 2 years ago

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。

bib_20211024 00 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