issues
search
AtsukiOsanai
/
cv_survey
Personal repository for computer vision survey
2
stars
0
forks
source link
Road Obstacle Detection Method Based on an Autoencoder with Semantic Segmentation
#14
Closed
AtsukiOsanai
closed
3 years ago
AtsukiOsanai
commented
3 years ago
ACCV2020
Paper
AtsukiOsanai
commented
3 years ago
Method
AutoEncoder Part
Sem-SegにはICNetを利用
Resynthesizeには
cascaded refinement network
を利用
ただし,Softmaxの出力ではなくその1つ前の特徴マップを入力に使う
Encoder (=Sem-Seg)とDecoder (=cascaded refinement network)のみが3つのmoduleの中でlearnableなmodule
Anormaly calculator
未知物体に対する仮説
未知物体の周辺には高い曖昧性が発生する (Sem-SegラベルのEntropyで評価)
曖昧な特徴マップから生成したResynthesized Imageはオリジナルの画像との差分が大きい
(VGG19のPerceptual Lossで計算)
これらの積を異常スコアとして定義
Post Processer
unknown obstacleのregionを抜くのが目的
まず画像をsuperpixelに分解
superpixelが障害物領域をまたぐことはほとんどないので精度向上に貢献する
最後にobstacle scoreを計算する
α_i: superpixel i内のanormaly scoreの平均値
n_j: superpixel jに含まれる画素の総数
p_j: superpixel jのroad labelの確率の平均値
r_ij: superpixel i, jの中心間距離
w: superpixel間の距離の中間値
L_iがthreshより高いsuperpixelをunknown obstacleと判定する
AtsukiOsanai
commented
3 years ago
Experiment on Lost and Found dataset
ROC curveによる解析
3種のベースラインよりも良好
ego-vehicleを除いた全てのクラスを対象にして場合はpost processingが効く
可視化
Ablation study
EntropyもPerceptual Lossも精度に寄与している
AtsukiOsanai
commented
3 years ago
Experiment on Highway Anomaly dataset
validation setを用いてL_iのthresholdを決定する
可視化による評価
ダイレクトにanomalyの領域を認識する方法との比較
特にpost-processingの効果が高い