Project 3: Fast R-CNN for Object Detection
1. Objectives. Recently, deep CNN have significantly improved image classification and object detection accuracy. Compared to image classification, object detection is a more challenging task that requires more complex methods to solve.
Example 1: single-class
Example 2: multi-classes
In this project, we learn a popular model for object detection, Fast R-CNN. You are required to realize object detecion using Fast R-CNN model and evaluate the results numerically. 2. The project includes the following two parts. 2.1 Single-class object detection in one image Use pre-trained Fast R-CNN model to do object detection in one example image. You need to carefully select the threshold of probability that a RoI is accepted as a detection. Only output the detection of car class.Plot the number of detections in the image over the value of threshold. Report your finally chosen threshold. Visualize the detected bounding boxes and the corresponding probability score of every bounding box in the image. 2.2 Object detection on Pascal VOC 2007 dataset Use the same model to do object detection on testing dataset of Pascal VOC 2007. There are 20 classes in the dataset. Use the same threshold as you choose in the first part. For each detection in a image, we compare it with ground truth annotations of the images. If there exists an annotation which has an (> 50%) overlap with the detection, we define the detection as a true positive. Show one example that contains true positive detections of multi-classes. To quantitatively evaluate the detection results, plot the precision-recall curve for car class. Report the average precision of every catego- ry and calculate the mean average precision (MAP) over the 20 classes.
3. Code and data Code and data can be downloaded here. For full instructions, please download this pdf file project description. 4. References  Girshick, Ross. "Fast r-cnn." Proceedings of the IEEE international conference on computer vision. 2015.[pdf]