How can I properly run Detectron2 on Videos?

Question

I'm currently working with Detectron2 for people detectios in videos, I've been trying to run the following code to read a video file, make the prediction frame by frame and record a video with the processed frames, but I am getting an empty video file. The environment that I've created for this is located in Colaboratory and has the following versions (python 3.6 , opencv 4.2.30). I'm new on this but I would truly appreciate if someone could give me an idea

This is the code

#!/usr/bin/env python3
# -- coding: utf-8 --

import detectron2
from detectron2.utils.logger import setup_logger
setup_logger()
# import some common libraries
import numpy as np
import cv2
# import some common detectron2 utilities
from detectron2 import model_zoo
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg
from detectron2.utils.visualizer import Visualizer
from detectron2.data import MetadataCatalog
import time

cap = cv2.VideoCapture('piloto legger 1.mp4')
hasFrame, frame = cap.read()
FPS = cap.get(cv2.CAP_PROP_FPS)
frame_width = frame.shape[1]
frame_height = frame.shape[0]
video_writer = cv2.VideoWriter('out.mp4', cv2.VideoWriter_fourcc('m', 'p', '4', 'v'), FPS, (frame_width, frame_height))
while cv2.waitKey(1) < 0:
    hasFrame, frame = cap.read()
    if not hasFrame:
        cv2.waitKey()
        break
    cfg = get_cfg()
    cfg.merge_from_file(model_zoo.get_config_file("COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x.yaml"))
    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.7  # set threshold for this model
    cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x.yaml")
    predictor = DefaultPredictor(cfg)
    outputs = predictor(frame)
    v = Visualizer(frame[:,:,::-1], MetadataCatalog.get(cfg.DATASETS.TRAIN[0]), scale=1.2)
    v = v.draw_instance_predictions(outputs["instances"].to("cpu"))
    imagen = (v.get_image()[:, :, ::-1])
    cv2.imwrite('POSE detectron2.png', imagen)
    video_writer.write(imagen)

cap.release()
video_writer.release()
cv2.destroyAllWindows()

Answer 1

I used your code as a starting point, and took some ideas from the Detectron2 examples in order to make it work.

The problem seems to have been something with the fourcc-argument of the VideoWriter, but may also have been related to your code using Visualizer instead of VideoVisualizer (and with a scale of 1.2, which made the image the wrong size for the VideoWriter).

The code below works for me (and is also a lot faster, as the predictor and visualizer are defined outside of the loop):

#!/usr/bin/env python3
# -- coding: utf-8 --

import detectron2
from detectron2.utils.logger import setup_logger
setup_logger()
# import some common libraries
import numpy as np
import tqdm
import cv2
# import some common detectron2 utilities
from detectron2 import model_zoo
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg
from detectron2.utils.video_visualizer import VideoVisualizer
from detectron2.utils.visualizer import ColorMode, Visualizer
from detectron2.data import MetadataCatalog
import time

# Extract video properties
video = cv2.VideoCapture('video-input.mp4')
width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT))
frames_per_second = video.get(cv2.CAP_PROP_FPS)
num_frames = int(video.get(cv2.CAP_PROP_FRAME_COUNT))

# Initialize video writer
video_writer = cv2.VideoWriter('out.mp4', fourcc=cv2.VideoWriter_fourcc(*"mp4v"), fps=float(frames_per_second), frameSize=(width, height), isColor=True)

# Initialize predictor
cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.7  # set threshold for this model
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml")
predictor = DefaultPredictor(cfg)

# Initialize visualizer
v = VideoVisualizer(MetadataCatalog.get(cfg.DATASETS.TRAIN[0]), ColorMode.IMAGE)

def runOnVideo(video, maxFrames):
    """ Runs the predictor on every frame in the video (unless maxFrames is given),
    and returns the frame with the predictions drawn.
    """

    readFrames = 0
    while True:
        hasFrame, frame = video.read()
        if not hasFrame:
            break

        # Get prediction results for this frame
        outputs = predictor(frame)

        # Make sure the frame is colored
        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)

        # Draw a visualization of the predictions using the video visualizer
        visualization = v.draw_instance_predictions(frame, outputs["instances"].to("cpu"))

        # Convert Matplotlib RGB format to OpenCV BGR format
        visualization = cv2.cvtColor(visualization.get_image(), cv2.COLOR_RGB2BGR)

        yield visualization

        readFrames += 1
        if readFrames > maxFrames:
            break

# Create a cut-off for debugging
num_frames = 120

# Enumerate the frames of the video
for visualization in tqdm.tqdm(runOnVideo(video, num_frames), total=num_frames):

    # Write test image
    cv2.imwrite('POSE detectron2.png', visualization)

    # Write to video file
    video_writer.write(visualization)

# Release resources
video.release()
video_writer.release()
cv2.destroyAllWindows()