I am using dlip package for face detection but it's too slow , and there is no significant improvement by turning GPU on.\nIs there any way to speed up this process? ( without reducing image size as it affect the detection accuracy)
", "messages": [ { "id": "747386", "postDate": "02/16/2020 11:17:21", "content": "I am using dlip package for face detection but it's too slow , and there is no significant improvement by turning GPU on.\nIs there any way to speed up this process? ( without reducing image size as it affect the detection accuracy)
", "rawMarkdown": "I am using dlip package for face detection but it's too slow , and there is no significant improvement by turning GPU on.\nIs there any way to speed up this process? ( without reducing image size as it affect the detection accuracy)", "votes": null }, { "id": "747416", "postDate": "02/16/2020 12:00:58", "content": "The MTCNN model from https://github.com/timesler/facenet-pytorch works very well. I find I can reduce the resolution of the videos by up to 25% to speed up processing without a meaningful drop in accuracy. It also runs on the GPU.
", "rawMarkdown": "The MTCNN model from https://github.com/timesler/facenet-pytorch works very well. I find I can reduce the resolution of the videos by up to 25% to speed up processing without a meaningful drop in accuracy. It also runs on the GPU.", "votes": null }, { "id": "747940", "postDate": "02/17/2020 03:06:25", "content": "Since dlib doesn't use a huge amount of GPU memory, you can use multiprocessing to fully utilise the GPUs. I used 4 jobs on 2 GPUs (8 jobs in total) and processed 2 million boxes over a couple of days
", "rawMarkdown": "Since dlib doesn't use a huge amount of GPU memory, you can use multiprocessing to fully utilise the GPUs. I used 4 jobs on 2 GPUs (8 jobs in total) and processed 2 million boxes over a couple of days", "votes": null }, { "id": "747941", "postDate": "02/17/2020 03:07:38", "content": "Many thanks James, it is at least 4 times faster .
", "rawMarkdown": "Many thanks James, it is at least 4 times faster .", "votes": null }, { "id": "747950", "postDate": "02/17/2020 03:18:25", "content": "Many thanks for your reply , i will try your solution later as I found James advice is good enough for me at this stat
", "rawMarkdown": "Many thanks for your reply , i will try your solution later as I found James advice is good enough for me at this stat", "votes": null }, { "id": "749167", "postDate": "02/18/2020 11:39:15", "content": "@jamesphoward how do we change video resolution ?
", "rawMarkdown": "jamesphoward how do we change video resolution ?", "votes": null }, { "id": "749170", "postDate": "02/18/2020 11:44:34", "content": "If you read the video in frame-by-frame with openCV, you can do frame = cv2.resize(frame, (new_width, new_height)) as you are reading in the frames.
@jamesphoward Hi! New to this competition. If you don't mind disclosing. \"Reducing the resolution of the videos by up to 25\" means cv2.resize(int(W*.75), int(H*.75))? \n Also, MTCNN seems to output resized and warped images, not bboxs. You are just using these cropped and resized images?
\n\nThanks in advance\nI actually meant to 25% of the original ie cv2.resize(frame_in, (width*0.25, height*0.25)), though I actually use 50% of the original at the moment.
MTCNN will output bounding boxes and probabilities rather than the images if you use boxes, probs = mtcnn.detect(pil_frames, landmarks=False) where boxes is a list of length of number of frames, and each element is a list of faces found in the video, and each face is (left, top, right, bottom)
The MTCNN model from https://github.com/timesler/facenet-pytorch works very well. I find I can reduce the resolution of the videos by up to 25% to speed up processing without a meaningful drop in accuracy. It also runs on the GPU.
", "votes": null, "replies": [ { "id": 747941, "author_name": "bosbos", "author_url": "", "post_date": "02/17/2020 03:07:38", "content": "Many thanks James, it is at least 4 times faster .
", "votes": null, "replies": [] }, { "id": 754728, "author_name": "roguekk007", "author_url": "", "post_date": "02/24/2020 01:27:30", "content": "@jamesphoward Hi! New to this competition. If you don't mind disclosing. \"Reducing the resolution of the videos by up to 25\" means cv2.resize(int(W*.75), int(H*.75))? \n Also, MTCNN seems to output resized and warped images, not bboxs. You are just using these cropped and resized images?
\n\nThanks in advance\nI actually meant to 25% of the original ie cv2.resize(frame_in, (width*0.25, height*0.25)), though I actually use 50% of the original at the moment.
MTCNN will output bounding boxes and probabilities rather than the images if you use boxes, probs = mtcnn.detect(pil_frames, landmarks=False) where boxes is a list of length of number of frames, and each element is a list of faces found in the video, and each face is (left, top, right, bottom)
Since dlib doesn't use a huge amount of GPU memory, you can use multiprocessing to fully utilise the GPUs. I used 4 jobs on 2 GPUs (8 jobs in total) and processed 2 million boxes over a couple of days
", "votes": null, "replies": [ { "id": 747950, "author_name": "bosbos", "author_url": "", "post_date": "02/17/2020 03:18:25", "content": "Many thanks for your reply , i will try your solution later as I found James advice is good enough for me at this stat
", "votes": null, "replies": [] } ] }, { "id": 749167, "author_name": "jaideepvalani", "author_url": "", "post_date": "02/18/2020 11:39:15", "content": "@jamesphoward how do we change video resolution ?
", "votes": null, "replies": [ { "id": 749170, "author_name": "jamesphoward", "author_url": "", "post_date": "02/18/2020 11:44:34", "content": "If you read the video in frame-by-frame with openCV, you can do frame = cv2.resize(frame, (new_width, new_height)) as you are reading in the frames.