refactor: use openfoodfacts.ml for object detection (#1493)

- use ml submodule for object detection
- add a ModelConfig pydantic class to specify model configuration.
  Previously, the model configuration was not clear and could lead to
  bugs.
- add a `threshold` parameter in model predict route

Dockerfile

@@ -5,7 +5,6 @@ ARG PYTHON_VERSION=3.11
 FROM python:$PYTHON_VERSION-slim AS python-base
 RUN apt-get update && \
     apt-get install --no-install-suggests --no-install-recommends -y gettext curl build-essential && \
-    apt-get install ffmpeg libsm6 libxext6  -y && \
     apt-get autoremove --purge && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*

pyproject.toml

@@ -72,9 +72,9 @@ python-redis-lock = "~4.0.0"
 transformers = "~4.44.2"
 lark = "~1.1.4"
 h5py = "~3.8.0"
-opencv-contrib-python-headless = "~4.10.0.84"
+opencv-python-headless = "~4.10.0.84"
 toml = "~0.10.2"
-openfoodfacts = "1.1.5"
+openfoodfacts = "2.3.4"
 imagehash = "~4.3.1"
 peewee-migrate = "~1.12.2"
 diskcache = "~5.6.3"

robotoff/app/api.py

@@ -875,6 +875,7 @@ class ImagePredictorResource:
     def on_get(self, req: falcon.Request, resp: falcon.Response):
         image_url = req.get_param("image_url", required=True)
         models: list[str] = req.get_param_as_list("models", required=True)
+        threshold: float = req.get_param_as_float("threshold", default=0.5)
 
         available_object_detection_models = list(
             ObjectDetectionModel.__members__.keys()
@@ -921,14 +922,16 @@ def on_get(self, req: falcon.Request, resp: falcon.Response):
                 model = ObjectDetectionModelRegistry.get(
                     ObjectDetectionModel[model_name]
                 )
-                result = model.detect_from_image(image, output_image=output_image)
+                result = model.detect_from_image(
+                    image, output_image=output_image, threshold=threshold
+                )
 
                 if output_image:
                     boxed_image = cast(Image.Image, result.boxed_image)
                     image_response(boxed_image, resp)
                     return
                 else:
-                    predictions[model_name] = result.to_json()
+                    predictions[model_name] = result.to_list()
             else:
                 model_enum = ImageClassificationModel[model_name]
                 classifier = image_classifier.ImageClassifier(

robotoff/cli/main.py

@@ -925,7 +925,7 @@ def import_logos(
     """
     from robotoff.cli import logos
     from robotoff.models import db
-    from robotoff.prediction.object_detection import OBJECT_DETECTION_MODEL_VERSION
+    from robotoff.prediction.object_detection import MODELS_CONFIG
     from robotoff.utils import get_logger
 
     logger = get_logger()
@@ -935,9 +935,7 @@ def import_logos(
         imported = logos.import_logos(
             data_path,
             ObjectDetectionModel.universal_logo_detector.value,
-            OBJECT_DETECTION_MODEL_VERSION[
-                ObjectDetectionModel.universal_logo_detector
-            ],
+            MODELS_CONFIG[ObjectDetectionModel.universal_logo_detector].model_version,
             batch_size,
             server_type,
         )

robotoff/insights/extraction.py

@@ -8,7 +8,7 @@
 from robotoff.off import get_source_from_url
 from robotoff.prediction import ocr
 from robotoff.prediction.object_detection import (
-    OBJECT_DETECTION_MODEL_VERSION,
+    MODELS_CONFIG,
     ObjectDetectionModelRegistry,
 )
 from robotoff.types import (
@@ -88,13 +88,13 @@ def run_object_detection_model(
     results = ObjectDetectionModelRegistry.get(model_name).detect_from_image(
         image, output_image=False, triton_uri=triton_uri, threshold=threshold
     )
-    data = results.to_json()
+    data = results.to_list()
     max_confidence = max((item["score"] for item in data), default=None)
     return ImagePrediction.create(
         image=image_model,
         type="object_detection",
         model_name=model_name.name,
-        model_version=OBJECT_DETECTION_MODEL_VERSION[model_name],
+        model_version=MODELS_CONFIG[model_name].model_version,
         data={"objects": data},
         timestamp=timestamp,
         max_confidence=max_confidence,

robotoff/prediction/object_detection/__init__.py

@@ -1,6 +1,2 @@
 # flake8: noqa
-from .core import (
-    OBJECT_DETECTION_MODEL_VERSION,
-    ObjectDetectionModelRegistry,
-    ObjectDetectionRawResult,
-)
+from .core import MODELS_CONFIG, ObjectDetectionModelRegistry, ObjectDetectionResult