Yahoo's NSFW model flags nudity; this labels the act

What it does

Miles Deep is a C++ application that classifies each second of a pornographic video into one of six categories of sexual act using a convolutional neural network with residual connections. It then edits the source video to keep only the requested acts, or strips out non-explicit scenes entirely. The author claims it is the first public tool to classify and edit pornography by specific act rather than merely detecting nudity.

The interesting bit

The project doubles as a general-purpose Caffe video-classification framework: by swapping the model weights, class definitions, and mean file, it can edit video for completely different subjects without recompiling. The included model is a ResNet pre-trained on ImageNet and fine-tuned on 36,000 explicit frames, with the top three layers retrained to reach roughly 95% accuracy on held-out test images.

Key highlights

• Distinguishes six specific categories plus an “other” catch-all, rather than applying a single generic NSFW tag
• Processes video at roughly 0.6 seconds per minute of input on a mid-range GPU, using batched classification and threaded frame capture
• Can output a cut video or a metadata tag file with timestamps and confidence scores for each detected segment
• Pluggable Caffe pipeline designed to accept new weights, model definitions, and mean files for other domains
• GPLv3 licensed, including the trained model weights

Caveats

• The training dataset is limited to two heterosexual performers and is not publicly released; only the fine-tuned weights are available
• No temporal modeling: each frame is classified independently, though the author notes that multi-frame approaches have struggled with camera movement anyway
• Explicitly targets Ubuntu 16.04 and CUDA 8.0; Windows and macOS support is mentioned but unfinished or untested by the author

Verdict

Useful if you need a working C++ template for batched Caffe video classification with threading. Less useful if you want a plug-and-play modern editor, since the dependency stack and unfinished cross-platform support mean you will likely be maintaining it yourself.