aimh-lab / visione Goto Github PK

Videos pre-processed for use during feature extraction are kept in the collection folder (named temp-videos-for-video-extraction)
If not in development mode, remove them (maybe create them as temporary files)

We have to decide a License and provides a list of dependencies

These two classes that orchestrate feature extraction from frames and videos share most of the functionalities, and the former can be generalized in the latter.

abbassare la quantizzazione da 7 secondi a 3 secondi
Abbassare l’intervallo da21 a 12 secondi

This should simplify managing the indices that should be loaded when serving:

Something like:

• visione analyze --replace features-clip-laion
• visione index --replace lucene-index
• visione index --replace --from lucene-documents

https://github.com/aimh-lab/visione/blob/61e049ec0fc195c5f78c398d3f30e0f6369b2ec0/visioneCore/src/main/java/it/cnr/isti/visione/services/VBSService.java#L212C29-L212C29

Boolean doALADIN=queryObj.getParameters().get("textualMode").indexOf("al") >= 0;

should be

Boolean doALADIN=queryObj.getParameters().get("textualMode").indexOf("ala") >= 0;

as also "all" start with "al"

This in not a real issue as doALADIN should be True also when textualMode="aladin", but may worth to fix it

We should add a .convert('RGB') here or similar.

sarebbe bello avere l'import per tutti i video di una cartella ,

dopo il for va ggiunto
token = f'4wc{label}{label_count_generale_non_divisa_per_rete}{4}'
tokens.append(token)

NB la label_count_generale_non_divisa_per_rete non comprende gli hyperset

The upload of image is not working - to be solved also in VISIONE RAI

As of now, text extraction at query time is performed on GPU. Change to CPU

• GEM da URL
• MVK UI: manca la ricerca da url
• gestire il conteggio oggetti in maniera dinamica nella UI (llinea 208 di visione.js)

ffmpeg inside multiprocessing causes some unwanted deadlock (at random points during the extraction)

As of now, this is patched in 1c9d63d by setting parallel=False

PROs:

• its easier/faster to check&skip already existing results w.r.t. walking the filesystem
• no additional deps (built-in in python)
• still a multi-language self-contained file-based storage
• we can implement several logic on SQL (object filtering and counting, conditional indexing, etc.)

CONs:

• SQL-like management, we'll have to cope with migrations
• probably will use more disk space
• embeddings as BLOBs. Save and load are pretty fast though, on SSD:

  Bulk insertion of 100k 1024-d vectors took 5.6198 seconds.
  Reading out 100k 1024-d vectors took 2.3810 seconds.

Configurazione testata:
Docker version 20.10.21, build 20.10.21-0ubuntu1~20.04.2
docker-compose version 1.25.0
Python 3.8.10

All'esecuzione del comando "visione" con qualsiasi parametro, ricevo l'errore
File "visione", line 8, in
sys.exit(main.main())
File "python3.8/site-packages/visione/cli/main.py", line 12, in main
compose_dir = Path(importlib.resources.files("visione.services").joinpath('')) # directory containing services for docker compose
AttributeError: module 'importlib.resources' has no attribute 'files'