Results of opencv dnn looks weird when using the tensorflow model(faster rcnn) to detect the object

edit

Link of the model zoo-- download model zip

• OS : win 10 64
• opencv version : 4.1.2, installed by anaconda

Using the opencv dnn module to perform object detection by the tensorflow model. But the results seems weird.

Steps to reproduce

1 : Generate the config file by tf_text_graph_faster_rcnn.py

python tf_text_graph_faster_rcnn.py --input frozen_inference_graph.pb --config pipeline.config --output faster_rcnn_inception_resnet_v2_atrous_oid.pbtxt

frozen_inference_graph.pb and pipeline.config are the files after unzip

2 : detect object by the example codes

import cv2 as cv

cvNet = cv.dnn.readNetFromTensorflow('tensorflow/faster_rcnn_inception_resnet_v2_atrous_oid.pb', 'tensorflow/faster_rcnn_inception_resnet_v2_atrous_oid.pbtxt')

img = cv.imread('traffic_jam.jpg')
rows = img.shape[0]
cols = img.shape[1]
cvNet.setInput(cv.dnn.blobFromImage(img, size=(300, 300), swapRB=True, crop=False))
cvOut = cvNet.forward()

for detection in cvOut[0,0,:,:]:
    score = float(detection[2])
    if score > 0.1:
        left = detection[3] * cols
        top = detection[4] * rows
        right = detection[5] * cols
        bottom = detection[6] * rows
        cv.rectangle(img, (int(left), int(top)), (int(right), int(bottom)), (23, 230, 210), thickness=2)

cv.imshow('img', img)
cv.waitKey()

1. The results

Input image

results

Almost every cars cannot detected by this network, I have tried to change the size of image, but results become worse in this case. How should I fix this issue?Thanks

Edit : If I use the tensorflow api(1.15) to do the object detection(example on this page)the results seems much better, although there are many vehicles cannot detected.