Faster RCNN
An another pytorch implementation of Faster RCNN base on https://github.com/longcw/faster_rcnn_pytorch, with rewriten data pre-process module and a lot of helpful debug messages.
Document: https://faster-rcnn.readthedocs.io/en/latest/
Installation
- Install docker and Nvidia-docker
- Git clone
git clone https:github.com/anhlt/faster_rcnn
- Using
docker-compose
docker-compose up --build
Modules
Dataset module
Extend the torchvision.datasets.CocoDetection module, and add extracting bounding box function. Image data will be automatically resize, and normalize
import os
import torchvision.transforms as transforms
from faster_rcnn.utils.dataset import CocoData
from faster_rcnn.utils.data_generator import CocoGenerator
from faster_rcnn.utils.data_generator import Enqueuer
dataDir = './data/mscoco'
dataType = 'train2014'
annFile='%s/annotations/instances_%s.json'%(dataDir,dataType)
images_dir = os.path.join(dataDir,'images', dataType)
cap = CocoData(root = images_dir,
annFile = annFile,
)You can verify your dataset module with this method
def imshow(inp, gt_boxes=[], predict_boxes = []):
"""Imshow for Tensor."""
inp = inp.numpy().transpose((1, 2, 0))
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
inp = std * inp + mean
inp = np.clip(inp, 0, 1)
fig,ax = plt.subplots(1, figsize=(20, 10))
ax.imshow(inp)
for i, box in enumerate(gt_boxes):
rect = patches.Rectangle((box[0], box[1]), box[2] - box[0], box[3] - box[1] ,linewidth=2,edgecolor='r',facecolor='none')
# Add the patch to the Axes
ax.add_patch(rect)
for i, box in enumerate(predict_boxes):
rect = patches.Rectangle((box[0], box[1]), box[2] - box[0], box[3] - box[1] ,linewidth=1,edgecolor='g',facecolor='none')
# Add the patch to the Axes
ax.add_patch(rect)
data = cap[13499]
im = data['tensor']
gt_boxes = data['boxes']
imshow(im[0], gt_boxes)