Inference on Videos

[simurgh7]

Hi,
Thanks a lot for creating this repository.
I was trying to run inference on this code by using video data and modifying the scripts. However, after a few frames later it runs into error like this -
Traceback (most recent call last):
File "infer_video.py", line 168, in
main()
File "infer_video.py", line 111, in main
result = coco_demo.overlay_keypoints_graph(img, top_predictions,vis_color ,target='car')
File "/code/lib/predictor.py", line 317, in overlay_keypoints_graph
keypoints = predictions.get_field("keypoints")
File "/maskrcnn-benchmark/maskrcnn_benchmark/structures/bounding_box.py", line 43, in get_field
return self.extra_fields[field]
KeyError: 'keypoints'

The modified snippet looks like this -
cap = cv2.VideoCapture(url_list[0])
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(output_dir + "/"+ now+"out.avi", fourcc, 30.0, (640, 480))
count = 0
while(True):
ret, img = cap.read()
if ret:
predictions = coco_demo.compute_prediction(img)
top_predictions = coco_demo.select_top_predictions(predictions)
result = coco_demo.overlay_keypoints_graph(img, top_predictions,vis_color ,target='car')
out.write(result)
#cv2.imshow('result',result)
count += 1
else:
break
cap.release()
out.release()
print(now)
now = datetime.datetime.now().strftime("%Y-%m-%d%H-%M-%S")
print(now)

Any idea how to get around this one? Thank you.

[dineshreddy91]

this might be an issue with no bounding box detected in the video frame. I would suggest adding a check to verify if the frame in the video has any detection if not just to continue.