add video support to segment-anything-2 pipeline #181
Conversation
| labels=kwargs.get('point_labels', None), | ||
| ) | ||
| video_segments = {} # video_segments contains the per-frame segmentation results | ||
| for out_frame_idx, out_obj_ids, out_mask_logits in self.tm_vid.propagate_in_video(inference_state): |
There was a problem hiding this comment.
I have a feeling we should return the full triple instead of creating a video segment, leaving post-processing to the consumer of the API ( though I recognize this is good quick way to validate the sanity of the mask outputs )
There was a problem hiding this comment.
Thanks for the guidance on this, I see how we can just return the results of self.tm_vid.propagate_in_video(inference_state):
There was a problem hiding this comment.
I have a feeling we should return the full triple instead of creating a video segment, leaving post-processing to the consumer of the API ( though I recognize this is good quick way to validate the sanity of the mask outputs )
I had some issues trying to return the correct values. I added frame index as an input parameter, normally propagate_in_video will loop returning results for each frame starting at the frame index until the end of the video, now it should only return a single frame. But the data doesn't look correct, can you take a look? @pschroedl
eliteprox
force-pushed
the
feature/sam2-add-video
branch
from
f8a45d9
to
3040cbd
Compare
eliteprox
changed the base branch from
segment_anything_2_pipeline_image
to
main
This change updates the request field
imagetomedia_fileand loads the appropriate segment-anything-2 inference model based on the content type of the file. It uses ffmpeg to process the video to image frames and loads them in with inference values from the request.Some adjustments still need to be made to the request/response parameters. I think "frame index" and "object id' may be two request parameters to add to this pipeline for the video requests, I've hard coded some values for now.