From this paper (section 3.2.1, emphasize mine):
\n\n\n", "messages": [ { "id": "765207", "postDate": "03/06/2020 11:26:43", "content": "Based on the observation that temporal coherence is not enforced effectively in the synthesis process of\n deepfakes, Sabir et al. [74] leveraged the use of spatio-temporal features of video streams to detect deepfakes.\n Video manipulation is carried out on a frame-by-frame basis so that low level artifacts produced by face\n manipulations are believed to further manifest themselves as temporal artifacts with inconsistencies across\n frames. A recurrent convolutional model (RCN) was proposed based on the integration of the convolutional\n network DenseNet [58] and the gated recurrent unit cells [75] to exploit temporal discrepancies across frames\n (see Fig. 3). The proposed method is tested on the FaceForensics++ data set, which includes 1,000 videos\n [76], and shows promising results
\n
From this paper (section 3.2.1, emphasize mine):
\n\n\n", "rawMarkdown": "From this [paper ](https://arxiv.org/pdf/1909.11573.pdf)(section 3.2.1, emphasize mine):\n\n> Based on the observation that **temporal coherence is not enforced effectively in the synthesis process of\ndeepfakes**, Sabir et al. [74] leveraged the use of spatio-temporal features of video streams to detect deepfakes.\nVideo manipulation is carried out on a frame-by-frame basis so that **low level artifacts produced by face\nmanipulations are believed to further manifest themselves as temporal artifacts** with inconsistencies across\nframes. A recurrent convolutional model (RCN) was proposed based on the integration of the convolutional\nnetwork DenseNet [58] and the gated recurrent unit cells [75] to exploit temporal discrepancies across frames\n(see Fig. 3). The proposed method is tested on the FaceForensics++ data set, which includes 1,000 videos\n[76], and shows promising results", "votes": null }, { "id": "765279", "postDate": "03/06/2020 12:57:59", "content": "Based on the observation that temporal coherence is not enforced effectively in the synthesis process of\n deepfakes, Sabir et al. [74] leveraged the use of spatio-temporal features of video streams to detect deepfakes.\n Video manipulation is carried out on a frame-by-frame basis so that low level artifacts produced by face\n manipulations are believed to further manifest themselves as temporal artifacts with inconsistencies across\n frames. A recurrent convolutional model (RCN) was proposed based on the integration of the convolutional\n network DenseNet [58] and the gated recurrent unit cells [75] to exploit temporal discrepancies across frames\n (see Fig. 3). The proposed method is tested on the FaceForensics++ data set, which includes 1,000 videos\n [76], and shows promising results
\n
I believe @jamesphoward is using EFNB3 with temporal layers (can't remember if it was an RNN or LSTM) and it seems to be fruitful ...
", "rawMarkdown": "I believe @jamesphoward is using EFNB3 with temporal layers (can't remember if it was an RNN or LSTM) and it seems to be fruitful ...", "votes": null }, { "id": "765933", "postDate": "03/07/2020 11:44:40", "content": "Thanks for the reference! Will give it a try.
", "rawMarkdown": "Thanks for the reference! Will give it a try.", "votes": null }, { "id": "765934", "postDate": "03/07/2020 11:45:01", "content": "One interesting blog post I stumbled upon: https://www.dessa.com/post/deepfake-detection-that-actually-works
", "rawMarkdown": "One interesting blog post I stumbled upon: https://www.dessa.com/post/deepfake-detection-that-actually-works", "votes": null } ], "comments": [ { "id": 765279, "author_name": "ma7moud", "author_url": "", "post_date": "03/06/2020 12:57:59", "content": "I believe @jamesphoward is using EFNB3 with temporal layers (can't remember if it was an RNN or LSTM) and it seems to be fruitful ...
", "votes": null, "replies": [ { "id": 765933, "author_name": "yassinealouini", "author_url": "", "post_date": "03/07/2020 11:44:40", "content": "Thanks for the reference! Will give it a try.
", "votes": null, "replies": [] } ] }, { "id": 765934, "author_name": "yassinealouini", "author_url": "", "post_date": "03/07/2020 11:45:01", "content": "One interesting blog post I stumbled upon: https://www.dessa.com/post/deepfake-detection-that-actually-works
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "765207": "From this [paper ](https://arxiv.org/pdf/1909.11573.pdf)(section 3.2.1, emphasize mine):\n\n> Based on the observation that **temporal coherence is not enforced effectively in the synthesis process of\ndeepfakes**, Sabir et al. [74] leveraged the use of spatio-temporal features of video streams to detect deepfakes.\nVideo manipulation is carried out on a frame-by-frame basis so that **low level artifacts produced by face\nmanipulations are believed to further manifest themselves as temporal artifacts** with inconsistencies across\nframes. A recurrent convolutional model (RCN) was proposed based on the integration of the convolutional\nnetwork DenseNet [58] and the gated recurrent unit cells [75] to exploit temporal discrepancies across frames\n(see Fig. 3). The proposed method is tested on the FaceForensics++ data set, which includes 1,000 videos\n[76], and shows promising results", "765279": "I believe @jamesphoward is using EFNB3 with temporal layers (can't remember if it was an RNN or LSTM) and it seems to be fruitful ...", "765933": "Thanks for the reference! Will give it a try.", "765934": "One interesting blog post I stumbled upon: https://www.dessa.com/post/deepfake-detection-that-actually-works" }, "source": "meta" }