Productionize Soccer Stream Detection

[cyberj0g]

What does this pull request do? Explain your changes. (required)
Make content detection functionality production-ready.

Limitations
Current implementation de facto uses a separate ad-hoc data stream to output results - the webhook. Detection results are not included with the renditions, because the DNN filter is run in a separate output, which doesn't produce any renditions. Model output is included into AVFrame metadata only for the purpose of transferring data back to LPMS. This would work for internal process of unwanted content detection, but the process needs to be re-worked to produce 'output stream' to end user, if required in the future for ML capabilities.

Specific updates (required)

• Add -detectContent command line flag to enable content detection for transcoder
• Fix list of T capabilities to include SceneDetection capability, if enabled
• Create correct list of T capabilities in case capabilities testing is disabled

How did you test each of these updates (required)

• Stdout echo web server for content detection webhook

while true; do {   echo -ne "HTTP/1.0 200 OK\r\nContent-Length: 0"; } | nc -l -p 8081 -q 1 ; done

• B

livepeer -broadcaster -rtmpAddr :1936 -cliAddr 127.0.0.1:7937 -nvidia all -httpAddr 127.0.0.1:8936 -orchAddr 127.0.0.1:8935 -orchSecret 11111111111 -v 6 -detectionWebhookUrl http://localhost:8081/

• OT

livepeer -transcoder -orchestrator -serviceAddr 127.0.0.1:8935 -nvidia all -v 7 -detectContent

• Stream

ffmpeg -re -i bbb.mp4 -c:a copy -c:v copy -f hls -hls_time 2 -headers 'Livepeer-Transcode-Configuration:
                                                                    {"detection": {"freq": 2, "sampleRate": 10, "sceneClassification": [{"name": "soccer"}]}}' http://localhost:8936/live/movie3/

• Webhook output

{"manifestID":"movie3","seqNo":10,"sceneClassification":[{"name":"soccer","probability":0.00676799938082695}]}

Does this pull request close any open issues?

#2394

Checklist:

• Read the contribution guide
• make runs successfully
• All tests in ./test.sh pass
• [n/a] README and other documentation updated
• Pending changelog updated

TBD: more tests for standalone T

[codecov]

Codecov Report

Merging #2431 (ddc9c1e) into master (d360781) will decrease coverage by 0.11113%.
The diff coverage is 11.36364%.

@@                 Coverage Diff                 @@
##              master       #2431         +/-   ##
===================================================
- Coverage   55.02775%   54.91662%   -0.11113%     
===================================================
  Files             94          94                 
  Lines          19641       19668         +27     
===================================================
- Hits           10808       10801          -7     
- Misses          8236        8271         +35     
+ Partials         597         596          -1     

Impacted Files	Coverage Δ
cmd/livepeer/starter/starter.go	2.36220% <0.00000%> (-0.01405%)	⬇️
server/ot_rpc.go	25.21994% <17.39130%> (-0.63676%)	⬇️
cmd/livepeer/livepeer.go	46.62921% <100.00000%> (+0.30152%)	⬆️
common/util.go	61.58537% <0.00000%> (-2.43902%)	⬇️
discovery/db_discovery.go	67.01389% <0.00000%> (-1.04167%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d360781...ddc9c1e. Read the comment docs.

[cyberj0g]

Converting this to draft, because discovered the following while testing:

1. Doesn't work in a separate O-T configuration, because detection results are not transmitted between T and O by HTTP
2. Detection webhook, which I expect to be the primary way of detecting unwanted content, not get called when verification is enabled.

[thomshutt]

Looks good, do we have any metrics around this at the moment? It'd be good when we begin testing to be able to see how many detections we're getting

[cyberj0g]

Looks good, do we have any metrics around this at the moment? It'd be good when we begin testing to be able to see how many detections we're getting

No, there are no metrics there. I expect that classification results will be reported through webhook, and metrics probably should be logged on that end.

[cyberj0g]

@red-0ne what is the purpose of this test? It always fails, if Nvidia GPU is present.

[red-0ne]

You are right. I think it should setup its GetGPU/GetCPU
The idea was to fail when no GPUs or appropriate PCI devices are detected (it fails now if you have a real one)
A minimal resolution would be

	getGPU = func() ([]*gpu.GraphicsCard, error) {
		return []*gpu.GraphicsCard{}, nil
	}
	getPCI = func() ([]*pci.Device, error) {
		return []*pci.Device{}, nil
	}