song630 / traffic-sign-detection-learning

A Faster-RCNN Approach

Traffic-Sign-Detection-Learning

---

Further modifications on this repo will be updated on Gitlab temporarily for about a month (before 9.15).

1. Choose Faster RCNN or Mask RCNN (rewrite and run)
2. Understand RoIAlign & RoIPooling
3. Read two papers:

Co-domain Embedding using Deep Quadruplet Networks for Unseen Traffic Sign Recognition

Perceptual Generative Adversarial Networks for Small Object Detection

Under Directory Test1 - Test3:

1. Three DL programs to quickly warm up and get familiar with PyTorch; more importantly, to gain experience on how to adjust parameters and architecture.
2. Test1: A simple FCN implementation on PASCAL-VOC-2007.
3. Test2: A classification network using CNN & ResNet on CIFAR-10.
4. Test3: A classfication & localization network on CUB-200-2011.

Under Directory Faster-RCNN-records:

1. Do multiple tests on PASCAL-VOC-2007 with Faster-RCNN from faster-rcnn.pytorch.
2. Records all the runtime output.

Under Directory tt100k-records:

1. Do tests based on the same source code.
2. Modify the pretreatment and evaluation parts of the code.
3. Records all the runtime output.

Under Directory cloned_faster_rcnn_codes:

1. Comment on the Faster-RCNN source code.

Under Directory pooling:

1. Read and comment on the code of RoIAlign & RoIPooling.

About the Notes:

1. note1.txt: The progress of object detection, starts from RCNN, through SPP and Fast-RCNN, ends with Faster-RCNN.
2. note2.txt: About RPN, etc.
3. note3.txt: Notes on Tutorial provided by faster-rcnn.pytorch.
4. note4.txt: Details on RoI-Pooling, RoI-Align and RoI-Warping.
5. note5.txt: Study PASCAL-VOC-2007 & TT100K datasets.

traffic-sign-detection-learning's Issues

train tt100k dataset but get nan losses

HI ,when I train tt100k I get bbox_loss=0 and then I get nan loss, I don't know if the dataset label was right,can you help me?