add video support

tests/integ/test_tools.py

@@ -11,7 +11,8 @@
  dpt_hybrid_midas,
  florence2_image_caption,
  florence2_ocr,
- florence2_phrase_grounding,
+ florence2_phrase_grounding_image,
+ florence2_phrase_grounding_video,
  florence2_roberta_vqa,
  florence2_sam2_image,
  florence2_sam2_video_tracking,
@@ -92,19 +93,19 @@ def test_owl_v2_video():
  assert 24 <= len([res["label"] for res in result[0]]) <= 26
 
 
-def test_florence2_phrase_grounding():
+def test_florence2_phrase_grounding_image():
  img = ski.data.coins()
- result = florence2_phrase_grounding(
+ result = florence2_phrase_grounding_image(
  image=img,
  prompt="coin",
  )
  assert len(result) == 25
  assert [res["label"] for res in result] == ["coin"] * 25
 
 
-def test_florence2_phrase_grounding_fine_tune_id():
+def test_florence2_phrase_grounding_image_fine_tune_id():
  img = ski.data.coins()
- result = florence2_phrase_grounding(
+ result = florence2_phrase_grounding_image(
  prompt="coin",
  image=img,
  fine_tune_id=FINE_TUNE_ID,
@@ -114,6 +115,32 @@ def test_florence2_phrase_grounding_fine_tune_id():
  assert [res["label"] for res in result] == ["coin"] * len(result)
 
 
+def test_florence2_phrase_grounding_video():
+ frames = [
+ np.array(Image.fromarray(ski.data.coins()).convert("RGB")) for _ in range(10)
+ ]
+ result = florence2_phrase_grounding_video(
+ prompt="coin",
+ frames=frames,
+ )
+ assert len(result) == 10
+ assert 24 <= len([res["label"] for res in result[0]]) <= 26
+
+
+def test_florence2_phrase_grounding_video_fine_tune_id():
+ frames = [
+ np.array(Image.fromarray(ski.data.coins()).convert("RGB")) for _ in range(10)
+ ]
+ # this calls a fine-tuned florence2 model which is going to be worse at this task
+ result = florence2_phrase_grounding_video(
+ prompt="coin",
+ frames=frames,
+ fine_tune_id=FINE_TUNE_ID,
+ )
+ assert len(result) == 10
+ assert 24 <= len([res["label"] for res in result[0]]) <= 26
+
+
 def test_template_match():
  img = ski.data.coins()
  result = template_match(

tests/unit/test_meta_tools.py

@@ -33,41 +33,41 @@ def test_use_object_detection_fine_tuning_none():
 
 def test_use_object_detection_fine_tuning():
  artifacts = Artifacts("test")
- code = """florence2_phrase_grounding('one', image1)
+ code = """florence2_phrase_grounding_image('one', image1)
 owl_v2_image('two', image2)
 florence2_sam2_image('three', image3)"""
- expected_code = """florence2_phrase_grounding("one", image1, "123")
+ expected_code = """florence2_phrase_grounding_image("one", image1, "123")
 owl_v2_image("two", image2, "123")
 florence2_sam2_image("three", image3, "123")"""
  artifacts["code"] = code
 
  output = use_object_detection_fine_tuning(artifacts, "code", "123")
- assert 'florence2_phrase_grounding("one", image1, "123")' in output
+ assert 'florence2_phrase_grounding_image("one", image1, "123")' in output
  assert 'owl_v2_image("two", image2, "123")' in output
  assert 'florence2_sam2_image("three", image3, "123")' in output
  assert artifacts["code"] == expected_code
 
 
 def test_use_object_detection_fine_tuning_twice():
  artifacts = Artifacts("test")
- code = """florence2_phrase_grounding('one', image1)
+ code = """florence2_phrase_grounding_image('one', image1)
 owl_v2_image('two', image2)
 florence2_sam2_image('three', image3)"""
- expected_code1 = """florence2_phrase_grounding("one", image1, "123")
+ expected_code1 = """florence2_phrase_grounding_image("one", image1, "123")
 owl_v2_image("two", image2, "123")
 florence2_sam2_image("three", image3, "123")"""
- expected_code2 = """florence2_phrase_grounding("one", image1, "456")
+ expected_code2 = """florence2_phrase_grounding_image("one", image1, "456")
 owl_v2_image("two", image2, "456")
 florence2_sam2_image("three", image3, "456")"""
  artifacts["code"] = code
  output = use_object_detection_fine_tuning(artifacts, "code", "123")
- assert 'florence2_phrase_grounding("one", image1, "123")' in output
+ assert 'florence2_phrase_grounding_image("one", image1, "123")' in output
  assert 'owl_v2_image("two", image2, "123")' in output
  assert 'florence2_sam2_image("three", image3, "123")' in output
  assert artifacts["code"] == expected_code1
 
  output = use_object_detection_fine_tuning(artifacts, "code", "456")
- assert 'florence2_phrase_grounding("one", image1, "456")' in output
+ assert 'florence2_phrase_grounding_image("one", image1, "456")' in output
  assert 'owl_v2_image("two", image2, "456")' in output
  assert 'florence2_sam2_image("three", image3, "456")' in output
  assert artifacts["code"] == expected_code2

vision_agent/agent/vision_agent_coder_prompts.py

@@ -101,15 +101,15 @@
 - Use the 'owl_v2_video' tool with the prompt 'person' to detect where the people are in the video.
 plan2:
 - Extract frames from 'video.mp4' at 10 FPS using the 'extract_frames_and_timestamps' tool.
-- Use the 'florence2_phrase_grounding' tool with the prompt 'person' to detect where the people are in the video.
+- Use the 'florence2_phrase_grounding_image' tool with the prompt 'person' to detect where the people are in the video.
 plan3:
 - Extract frames from 'video.mp4' at 10 FPS using the 'extract_frames_and_timestamps' tool.
 - Use the 'florence2_sam2_video_tracking' tool with the prompt 'person' to detect where the people are in the video.
 
 
 ```python
 import numpy as np
-from vision_agent.tools import extract_frames_and_timestamps, owl_v2_video, florence2_phrase_grounding, florence2_sam2_video_tracking
+from vision_agent.tools import extract_frames_and_timestamps, owl_v2_video, florence2_phrase_grounding_image, florence2_sam2_video_tracking
 
 # sample at 1 FPS and use the first 10 frames to reduce processing time
 frames = extract_frames_and_timestamps("video.mp4", 1)
@@ -143,7 +143,7 @@ def get_counts(preds):
 owl_v2_counts = get_counts(owl_v2_out)
 
 # plan2
-florence2_out = [florence2_phrase_grounding("person", f) for f in frames]
+florence2_out = [florence2_phrase_grounding_image("person", f) for f in frames]
 florence2_counts = get_counts(florence2_out)
 
 # plan3
@@ -153,13 +153,13 @@ def get_counts(preds):
 
 final_out = {{
  "owl_v2_video": owl_v2_out,
- "florence2_phrase_grounding": florence2_out,
+ "florence2_phrase_grounding_image": florence2_out,
  "florence2_sam2_video_tracking": f2s2_out,
 }}
 
 counts = {{
  "owl_v2_video": owl_v2_counts,
- "florence2_phrase_grounding": florence2_counts,
+ "florence2_phrase_grounding_image": florence2_counts,
  "florence2_sam2_video_tracking": f2s2_counts,
 }}
 

vision_agent/agent/vision_agent_prompts.py

@@ -131,10 +131,10 @@
 
 OBSERVATION:
 [Artifact code.py]
-0|from vision_agent.tools import load_image, florence2_phrase_grounding, closest_box_distance, overlay_bounding_boxes, save_image
+0|from vision_agent.tools import load_image, florence2_phrase_grounding_image, closest_box_distance, overlay_bounding_boxes, save_image
 1|def count_workers_with_helmets(image_path: str, output_path: str):
 2| image = load_image(image_path)
-3| detections = florence2_phrase_grounding("worker, helmet", image)
+3| detections = florence2_phrase_grounding_image("worker, helmet", image)
 4| workers = [d for d in detections if d['label'] == 'worker']
 5| helmets = [d for d in detections if d['label'] == 'helmet']
 6| count = 0
@@ -166,18 +166,18 @@
 OBSERVATION:
 [Fine tuning id: 23b3b022-5ebf-4798-9373-20ef36429abf]
 
-AGENT: {"thoughts": "The model has finished fine tuning, I will now replace the original florence2_phrase_grounding call with the fine tuning id.", "response": "I will not update the code to use the fine tuned model. <execute_python>use_object_detection_fine_tuning(artifacts, "code.py", "23b3b022-5ebf-4798-9373-20ef36429abf")</execute_python>", "let_user_respond": false}
+AGENT: {"thoughts": "The model has finished fine tuning, I will now replace the original florence2_phrase_grounding_image call with the fine tuning id.", "response": "I will not update the code to use the fine tuned model. <execute_python>use_object_detection_fine_tuning(artifacts, "code.py", "23b3b022-5ebf-4798-9373-20ef36429abf")</execute_python>", "let_user_respond": false}
 
 OBSERVATION:
 [Artifact code.py edits]
 ---
 +++
 @@ -1,7 +1,7 @@
- from vision_agent.tools import load_image, florence2_phrase_grounding, closest_box_distance, overlay_bounding_boxes, save_image
+ from vision_agent.tools import load_image, florence2_phrase_grounding_image, closest_box_distance, overlay_bounding_boxes, save_image
  def count_workers_with_helmets(image_path: str, output_path: str):
  image = load_image(image_path)
-- detections = florence2_phrase_grounding("worker, helmet", image)
-+ detections = florence2_phrase_grounding("worker, helmet", image, "23b3b022-5ebf-4798-9373-20ef36429abf")
+- detections = florence2_phrase_grounding_image("worker, helmet", image)
++ detections = florence2_phrase_grounding_image("worker, helmet", image, "23b3b022-5ebf-4798-9373-20ef36429abf")
  workers = [d for d in detections if d['label'] == 'worker']
  helmets = [d for d in detections if d['label'] == 'helmet']
  count = 0
@@ -189,5 +189,5 @@ def count_workers_with_helmets(image_path: str, output_path: str):
 ----- stdout -----
 3
 
-AGENT: {"thoughts": "Three workers with helmets are detected, I will show this to the user and ask them if the result looks good.", "response": "I have updated the code using the fine tuned florence2_phrase_grounding model to count the workers wearing helmets in code.py and saved the visualization under 'workers_viz.png'.", "let_user_respond": true}
+AGENT: {"thoughts": "Three workers with helmets are detected, I will show this to the user and ask them if the result looks good.", "response": "I have updated the code using the fine tuned florence2_phrase_grounding_image model to count the workers wearing helmets in code.py and saved the visualization under 'workers_viz.png'.", "let_user_respond": true}
 """

vision_agent/tools/__init__.py

@@ -24,7 +24,8 @@
  extract_frames_and_timestamps,
  florence2_image_caption,
  florence2_ocr,
- florence2_phrase_grounding,
+ florence2_phrase_grounding_image,
+ florence2_phrase_grounding_video,
  florence2_roberta_vqa,
  florence2_sam2_image,
  florence2_sam2_video_tracking,

vision_agent/tools/meta_tools.py

@@ -665,8 +665,12 @@ def use_object_detection_fine_tuning(
 
  patterns_with_fine_tune_id = [
  (
- r'florence2_phrase_grounding\(\s*["\']([^"\']+)["\']\s*,\s*([^,]+)(?:,\s*["\'][^"\']+["\'])?\s*\)',
- lambda match: f'florence2_phrase_grounding("{match.group(1)}", {match.group(2)}, "{fine_tune_id}")',
+ r'florence2_phrase_grounding_image\(\s*["\']([^"\']+)["\']\s*,\s*([^,]+)(?:,\s*["\'][^"\']+["\'])?\s*\)',
+ lambda match: f'florence2_phrase_grounding_image("{match.group(1)}", {match.group(2)}, "{fine_tune_id}")',
+ ),
+ (
+ r'florence2_phrase_grounding_video\(\s*["\']([^"\']+)["\']\s*,\s*([^,]+)(?:,\s*["\'][^"\']+["\'])?\s*\)',
+ lambda match: f'florence2_phrase_grounding_video("{match.group(1)}", {match.group(2)}, "{fine_tune_id}")',
  ),
  (
  r'owl_v2_image\(\s*["\']([^"\']+)["\']\s*,\s*([^,]+)(?:,\s*["\'][^"\']+["\'])?\s*\)',

vision_agent/tools/tools.py

@@ -1141,16 +1141,13 @@ def florence2_image_caption(image: np.ndarray, detail_caption: bool = True) -> s
  return answer[task] # type: ignore
 
 
-# TODO: add video
-
-
-def florence2_phrase_grounding(
+def florence2_phrase_grounding_image(
  prompt: str, image: np.ndarray, fine_tune_id: Optional[str] = None
 ) -> List[Dict[str, Any]]:
- """'florence2_phrase_grounding' is a tool that can detect multiple
- objects given a text prompt which can be object names or caption. You
- can optionally separate the object names in the text with commas. It returns a list
- of bounding boxes with normalized coordinates, label names and associated
+ """'florence2_phrase_grounding_image' will run florence2 on a image. It can
+ detect multiple objects given a text prompt which can be object names or caption.
+ You can optionally separate the object names in the text with commas. It returns
+ a list of bounding boxes with normalized coordinates, label names and associated
  probability scores of 1.0.
 
  Parameters:
@@ -1168,7 +1165,7 @@ def florence2_phrase_grounding(
 
  Example
  -------
- >>> florence2_phrase_grounding('person looking at a coyote', image)
+ >>> florence2_phrase_grounding_image('person looking at a coyote', image)
  [
  {'score': 1.0, 'label': 'person', 'bbox': [0.1, 0.11, 0.35, 0.4]},
  {'score': 1.0, 'label': 'coyote', 'bbox': [0.34, 0.21, 0.85, 0.5},
@@ -1196,7 +1193,7 @@ def florence2_phrase_grounding(
  data,
  "florence2-ft",
  v2=True,
- metadata_payload={"function_name": "florence2_phrase_grounding"},
+ metadata_payload={"function_name": "florence2_phrase_grounding_image"},
  )
  # get the first frame
  detection = detections[0]
@@ -1205,7 +1202,7 @@ def florence2_phrase_grounding(
  "image": image_b64,
  "task": "<CAPTION_TO_PHRASE_GROUNDING>",
  "prompt": prompt,
- "function_name": "florence2_phrase_grounding",
+ "function_name": "florence2_phrase_grounding_image",
  }
  detections = send_inference_request(data, "florence2", v2=True)
  detection = detections["<CAPTION_TO_PHRASE_GROUNDING>"]
@@ -1222,6 +1219,90 @@ def florence2_phrase_grounding(
  return [bbox.model_dump() for bbox in return_data]
 
 
+def florence2_phrase_grounding_video(
+ prompt: str, frames: List[np.ndarray], fine_tune_id: Optional[str] = None
+) -> List[Dict[str, Any]]:
+ """'florence2_phrase_grounding_video' will run florence2 on each frame of a video.
+ It can detect multiple objects given a text prompt which can be object names or
+ caption. You can optionally separate the object names in the text with commas.
+ It returns a list of lists where each inner list contains bounding boxes with
+ normalized coordinates, label names and associated probability scores of 1.0.
+
+ Parameters:
+ prompt (str): The prompt to ground to the video.
+ frames (List[np.ndarray]): The list of frames to detect objects.
+ fine_tune_id (Optional[str]): If you have a fine-tuned model, you can pass the
+ fine-tuned model ID here to use it.
+
+ Returns:
+ List[List[Dict[str, Any]]]: A list of lists of dictionaries containing the score,
+ label, and bounding box of the detected objects with normalized coordinates
+ between 0 and 1 (xmin, ymin, xmax, ymax). xmin and ymin are the coordinates
+ of the top-left and xmax and ymax are the coordinates of the bottom-right of
+ the bounding box. The scores are always 1.0 and cannot be thresholded.
+
+ Example
+ -------
+ >>> florence2_phrase_grounding_video('person looking at a coyote', frames)
+ [
+ [
+ {'score': 1.0, 'label': 'person', 'bbox': [0.1, 0.11, 0.35, 0.4]},
+ {'score': 1.0, 'label': 'coyote', 'bbox': [0.34, 0.21, 0.85, 0.5},
+ ],
+ ...
+ ]
+ """
+ if len(frames) == 0:
+ raise ValueError("No frames provided")
+
+ image_size = frames[0].shape[:2]
+ buffer_bytes = frames_to_bytes(frames)
+ files = [("video", buffer_bytes)]
+
+ if fine_tune_id is not None:
+ landing_api = LandingPublicAPI()
+ status = landing_api.check_fine_tuning_job(UUID(fine_tune_id))
+ if status is not JobStatus.SUCCEEDED:
+ raise FineTuneModelIsNotReady(
+ f"Fine-tuned model {fine_tune_id} is not ready yet"
+ )
+
+ data_obj = Florence2FtRequest(
+ video=buffer_bytes,
+ task=PromptTask.PHRASE_GROUNDING,
+ prompt=prompt,
+ job_id=UUID(fine_tune_id),
+ )
+ data = data_obj.model_dump(by_alias=True, exclude_none=True)
+ else:
+ data_obj = Florence2FtRequest(
+ video=buffer_bytes, task=PromptTask.PHRASE_GROUNDING, prompt=prompt
+ )
+ data = data_obj.model_dump(by_alias=True, exclude_none=True)
+
+ detections = send_inference_request(
+ data,
+ "florence2-ft",
+ v2=True,
+ files=files,
+ metadata_payload={"function_name": "florence2_phrase_grounding_video"},
+ )
+
+ bboxes_formatted = []
+ for frame_data in detections:
+ bboxes_formatted_per_frame = []
+ for idx in range(len(frame_data["bboxes"])):
+ bboxes_formatted_per_frame.append(
+ ODResponseData(
+ label=frame_data["labels"][idx],
+ bbox=normalize_bbox(frame_data["bboxes"][idx], image_size),
+ score=1.0,
+ )
+ )
+ bboxes_formatted.append(bboxes_formatted_per_frame)
+ return [[bbox.model_dump() for bbox in frame] for frame in bboxes_formatted]
+
+
 def florence2_ocr(image: np.ndarray) -> List[Dict[str, Any]]:
  """'florence2_ocr' is a tool that can detect text and text regions in an image.
  Each text region contains one line of text. It returns a list of detected text,
@@ -1233,7 +1314,7 @@ def florence2_ocr(image: np.ndarray) -> List[Dict[str, Any]]:
 
  Returns:
  List[Dict[str, Any]]: A list of dictionaries containing the detected text, bbox
- with nornmalized coordinates, and confidence score.
+ with normalized coordinates, and confidence score.
 
  Example
  -------
@@ -2077,7 +2158,8 @@ def overlay_counting_results(
  florence2_ocr,
  florence2_sam2_image,
  florence2_sam2_video_tracking,
- florence2_phrase_grounding,
+ florence2_phrase_grounding_image,
+ florence2_phrase_grounding_video,
  ixc25_image_vqa,
  ixc25_video_vqa,
  detr_segmentation,