[논문 요약28] [11] R. Girshick. Fast r-cnn. In ICCV, pages 1440–1448, 2015. 1, 4, 5, 6, 7, 8

[업데이트 2018.09.07 10:15]

1. 베이스 논문

[논문 요약17] Perceptual Generative Adversarial Networks for Small Object Detection

(http://arclab.tistory.com/172)

Recent great progress on object detection is stimulated by the deep learning pipelines that learn deep representations from the region of interest (RoI) and perform classification based on the learned representations, such as Fast R-CNN [11] and Faster R-CNN [32].

The learned residual representation is then used to enhance the feature pooled from “Conv5” for the small object proposal through RoI pooling [11] by element-wise sum operation, producing super-resolved representation.

In particular, Lcls(p; g) = -log pg is log loss for the ground truth class g and Lloc is a smooth L1 loss proposed in [11].

Following [16], we perform down-sampling directly by convolutional layers with a stride of 2. The implementation is based on the publicly available Fast R-CNN framework [11] built on the Caffe platform [17].

Following [11], in each minibatch, 25% of object proposals are foreground that overlap with a ground truth bounding box with at least 0.5 IoU, and the rest are background.

“Multi-scale Input” indicates the model trained with input images with multi-scale settings(s 2 1120; 1340; 1600; 1920; 2300) as adopted in [11].

Our method achieves 69.4%, 60.2%, 57.9% and 41.8% in Average Precision (AP) for boat, bottle, chair, and plant, respectively. It significantly outperforms those of the Fast R-CNN [11] baseline, i.e., 59.4%, 38.3%, 42.8% and 31.8%, well demonstrating the generalization capability of the proposed generator for general small object detection.

2. 레퍼런스 논문

[11] R. Girshick. Fast r-cnn. In ICCV, pages 1440–1448, 2015. 1, 4, 5, 6, 7, 8

(https://arxiv.org/pdf/1504.08083.pdf)

3. 주요 내용 요약

3.1 레퍼런스 논문 주요 내용

Fast R-CNN에 대한 내용입니다.

(http://arclab.tistory.com/170)

3.2 베이스 논문에서 인용한 내용

본 논문에서는 ROI pooling 등 object detection을 위해 Fast R-CNN 아키텍쳐를 사용합니다.