Hey!
\n\nSince the dataset is huge and it'll be quite hard to source this information through code because of both compute times and false positives arising from face detection, I wanted to know if the number of faces in any given video remains the same throughout the video. Eg: if a video has only 1 face, then later in the video, can 1 or more faces enter/exit the video?
\n\nThanks,\nAkash
", "messages": [ { "id": "714576", "postDate": "01/09/2020 15:08:11", "content": "Hey!
\n\nSince the dataset is huge and it'll be quite hard to source this information through code because of both compute times and false positives arising from face detection, I wanted to know if the number of faces in any given video remains the same throughout the video. Eg: if a video has only 1 face, then later in the video, can 1 or more faces enter/exit the video?
\n\nThanks,\nAkash
", "rawMarkdown": "Hey!\n\nSince the dataset is huge and it'll be quite hard to source this information through code because of both compute times and false positives arising from face detection, I wanted to know if the number of faces in any given video remains the same throughout the video. Eg: if a video has only 1 face, then later in the video, can 1 or more faces enter/exit the video?\n\nThanks,\nAkash", "votes": null }, { "id": "714787", "postDate": "01/09/2020 18:41:55", "content": "Watched a bunch of the videos to get a feel for what kind of situations exist. I do recall one were two folks entered the scene with backs mostly facing the camera and than they turned to face it. Would guess that any face detection might see only one face after x number of frames and than see the other face a bit later.
\n\nSo my guess would be that answer to your question is Yes - but as an additional guess the number of files were this happens is small.
\n\nBecause of detection errors my current method does have a different count of faces during the 300 frames, and that's another much more important thing to fix for me.
", "rawMarkdown": "Watched a bunch of the videos to get a feel for what kind of situations exist. I do recall one were two folks entered the scene with backs mostly facing the camera and than they turned to face it. Would guess that any face detection might see only one face after x number of frames and than see the other face a bit later.\n\nSo my guess would be that answer to your question is Yes - but as an additional guess the number of files were this happens is small. \n\nBecause of detection errors my current method does have a different count of faces during the 300 frames, and that's another much more important thing to fix for me.", "votes": null }, { "id": "714815", "postDate": "01/09/2020 19:24:58", "content": "Ah! Thanks for the heads up!\nI agree with the different count of faces. One of the major problems I'm facing while using MTCNN is that Couches and Hands are being tagged as faces with confidence scores of >90%. So it'll be quite important to even find out the confidence threshold value below which we need to reject detection.
", "rawMarkdown": "Ah! Thanks for the heads up!\nI agree with the different count of faces. One of the major problems I'm facing while using MTCNN is that Couches and Hands are being tagged as faces with confidence scores of >90%. So it'll be quite important to even find out the confidence threshold value below which we need to reject detection.", "votes": null }, { "id": "714929", "postDate": "01/10/2020 00:02:15", "content": "There is a simple fastai kernel in one of their v3 sets of notebooks that does a GAN to create fake bedroom suite images. Maybe the data site being used has a bunch of couch images that can be used to train a model between faces and couches ?????
\n\nI think I am kidding but .....
", "rawMarkdown": "There is a simple fastai kernel in one of their v3 sets of notebooks that does a GAN to create fake bedroom suite images. Maybe the data site being used has a bunch of couch images that can be used to train a model between faces and couches ?????\n\nI think I am kidding but .....", "votes": null }, { "id": "746861", "postDate": "02/15/2020 16:40:00", "content": "found the same thing with the couches and hands lol 😂
", "rawMarkdown": "found the same thing with the couches and hands lol 😂", "votes": null } ], "comments": [ { "id": 714787, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2020 18:41:55", "content": "Watched a bunch of the videos to get a feel for what kind of situations exist. I do recall one were two folks entered the scene with backs mostly facing the camera and than they turned to face it. Would guess that any face detection might see only one face after x number of frames and than see the other face a bit later.
\n\nSo my guess would be that answer to your question is Yes - but as an additional guess the number of files were this happens is small.
\n\nBecause of detection errors my current method does have a different count of faces during the 300 frames, and that's another much more important thing to fix for me.
", "votes": null, "replies": [ { "id": 714815, "author_name": "akashnandi", "author_url": "", "post_date": "01/09/2020 19:24:58", "content": "Ah! Thanks for the heads up!\nI agree with the different count of faces. One of the major problems I'm facing while using MTCNN is that Couches and Hands are being tagged as faces with confidence scores of >90%. So it'll be quite important to even find out the confidence threshold value below which we need to reject detection.
", "votes": null, "replies": [] }, { "id": 714929, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/10/2020 00:02:15", "content": "There is a simple fastai kernel in one of their v3 sets of notebooks that does a GAN to create fake bedroom suite images. Maybe the data site being used has a bunch of couch images that can be used to train a model between faces and couches ?????
\n\nI think I am kidding but .....
", "votes": null, "replies": [] }, { "id": 746861, "author_name": "adamwlev", "author_url": "", "post_date": "02/15/2020 16:40:00", "content": "found the same thing with the couches and hands lol 😂
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "714576": "Hey!\n\nSince the dataset is huge and it'll be quite hard to source this information through code because of both compute times and false positives arising from face detection, I wanted to know if the number of faces in any given video remains the same throughout the video. Eg: if a video has only 1 face, then later in the video, can 1 or more faces enter/exit the video?\n\nThanks,\nAkash", "714787": "Watched a bunch of the videos to get a feel for what kind of situations exist. I do recall one were two folks entered the scene with backs mostly facing the camera and than they turned to face it. Would guess that any face detection might see only one face after x number of frames and than see the other face a bit later.\n\nSo my guess would be that answer to your question is Yes - but as an additional guess the number of files were this happens is small. \n\nBecause of detection errors my current method does have a different count of faces during the 300 frames, and that's another much more important thing to fix for me.", "714815": "Ah! Thanks for the heads up!\nI agree with the different count of faces. One of the major problems I'm facing while using MTCNN is that Couches and Hands are being tagged as faces with confidence scores of >90%. So it'll be quite important to even find out the confidence threshold value below which we need to reject detection.", "714929": "There is a simple fastai kernel in one of their v3 sets of notebooks that does a GAN to create fake bedroom suite images. Maybe the data site being used has a bunch of couch images that can be used to train a model between faces and couches ?????\n\nI think I am kidding but .....", "746861": "found the same thing with the couches and hands lol 😂" }, "source": "meta" }