Inference inputs

[lushihan]

Hello,

Thanks for sharing your work! Do you convert the video into the image sequence and then dump them to the model?

After you get the inferred images, you transform them into video again?

[trrahul]

I don't explicitly convert video into an image sequence, rather I use opencv's read() to read the video by frame and do inference on each of those frames iteratively.

The model outputs data points for one frame at a time, I plot the frame and the contours on a matplotlib canvas, and render the canvas it to an image file. After all frames are processed, I use ffmpeg to generate video using these image files.

[lushihan]

Thanks! Since I met some bugs following your instructions, do you put your input video file in the directory of your "infer_simply.py"?

[trrahul]

Yes, and you can put your video anywhere you want, just pass the location to cv2.VideoCapture().

There are a number of issues with the script.

1. There is a considerable overhead for file read/write operations as it writes each frame after inference. It can be solved if you keep all the processed frames in memory and write it once.
2. You need to know my setup or read through the code to use it.

I apologize for the clumsiness of the script but I intend to address these in the future.

Edit: I have updated the script. It shouldn't generate any more errors.