James Elsey

Random nuggets of usefulness

2 minute read

S3 Image generation not working?

In 2024, AWS announced a new feature for Kinesis Video Streams, which is a managed service where they can automatically take images from a video stream and store them into S3, without you having to write any code, perfect!

Having followed the documentation, I wasn’t able to get this working, after wasting hours, I was able to get it working, so I thought I’d save others some time by documenting what I did.

Data into KVS

It doesn’t really matter how you’re streaming video into KVS, but for the record I’m using the kvssink gstreamer plugin, from the Raspberry Pi. My pipeline looked like this, with the relevant part being the last line

gst-launch-1.0 shmsrc socket-path=/tmp/infered.feed do-timestamp=1 ! \
        video/x-raw,format=RGB,width=1920,height=1080,framerate=30/1 ! \
        queue max-size-buffers=2 leaky=downstream ! \
        videorate ! video/x-raw,framerate=30/1 ! \
        videoconvert ! video/x-raw,format=I420 ! \
        x264enc key-int-max=25 tune=zerolatency speed-preset=ultrafast bitrate=1000 ! \
        h264parse ! \
        kvssink stream-name="inferred" access-key="$AWS_ACCESS_KEY" secret-key="$AWS_SECRET_ACCESS_KEY" aws-region="$AWS_REGION" max-latency=100

KVS to S3

Following the documentation, you’ll create a file like this

{
 "StreamName": "TestStream",
 "ImageGenerationConfiguration":
 {
  "Status": "ENABLED",
  "DestinationConfig":
  {
   "DestinationRegion": "us-east-1",
   "Uri": "s3://bucket-name"
  },
  "SamplingInterval": 200,
  "ImageSelectorType": "PRODUCER_TIMESTAMP",
  "Format": "JPEG",
  "FormatConfig": {
                "JPEGQuality": "80"
       },
  "WidthPixels": 320,
  "HeightPixels": 240
 }
}

and then apply it using this

aws kinesisvideo update-image-generation-configuration \
--cli-input-json file://./update-image-generation-input.json \

Well, that still didn’t work for me, so I had to do some more digging.

Permissions?

As mentioned in this youtube video about image extraction, the user that kvssink is using (as in, the user the key & secret key belongs to) is the user that’s used when KVS tries to write to s3, so you need to make sure that has permissions on S3.

For demo purposes, I just gave it full access to S3, and I also updated the bucket policy to allow the kinesis service access.

Nope, still not working.

The fix

After some digging, I found that the KVS producer (in my case, kvssink), needs to send tags to the stream to tell it to use the image generation.

I tried to use the --stream-tags, as suggested elsewhere, but that didn’t work for me.

Then I stumbled on this PR to the KVS producer. This was it, this is what worked for me. I had to check out the fork and rebuild the SDK (same steps as in the README), and then I was able to add the tags to the stream by using the img-gen=1 on the kvssink

That PR was closed in favour of another one, but that still hasn’t been merged yet, so I’m not sure if this will be the final solution, but it worked for me.

Am I doing something wrong here? Why was this so difficult to setup, and the SDK still not updated to include this feature? Well, hopefully you don’t waste a bunch of hours like I did.

You May Also Enjoy