【目標檢測】小目標檢測問題及解決方法

本部分主要節選自 《Augmentation for small object detection》。
針對目標檢測中的小目標問題,主要有以下幾種解決方法:
(1) 增加輸入圖片分辨率 [1,2]
(2) 混合多尺度特徵 [3,4,5,6]
(3) 用 GAN 來區分大物體/小物體特徵,然後對小物體特徵轉化爲更精細的特徵以此加強小目標的檢測 [7]
(4) 檢測小物體時增加上下文信息 [8,9,10,11]
(5) 對包含小物體的圖片過採樣+單張圖片中多複製粘貼幾份小物體 [12]

後面均是使用相關方法的參考文獻,這裏給出它們的名字,有興趣的讀者可以自行搜索,拓展閱讀。

相關文獻:
[1] 3d object proposals for accurate object class detection.
[2] Ssd: Single shot multibox detector.
[3] Inside-outside net: Detecting objects in context with skip pooling and recurrent neural networks.
[4] Feature-fused ssd: fast detection for small objects.
[5] Cnn-based small object detection and visualization with feature activation mapping.
[6] Exploit all the layers: Fast and accurate cnn object detector with scale dependent pooling and cascaded rejection classifiers.
[7] Perceptual generative adversarial networks for small object detection.
[8] R-cnn for small object detection.
[9] Loco: Local context based faster r-cnn for small traffic sign detection.
[10] Finding tiny faces.
[11] Small object detection in optical remote sensing images via modified faster r-cnn.
[12] Augmentation for small object detection

發表評論
所有評論
還沒有人評論,想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章