| | import random |
| |
|
| | from osdsynth.utils.logger import SkipImageException |
| |
|
| | global_qs_list = [ |
| | "Can you provide a detailed description of this image in one paragraph of 30 words or less?", |
| | "Could you give a concise, detailed account of what's depicted in this image, all in one paragraph, aiming for no more than 30 words?", |
| | "Please offer a detailed portrayal of this image, condensed into one paragraph, keeping it under 30 words.", |
| | "Could you sketch out a detailed narrative of this image within a 30-word limit, all in a single paragraph?", |
| | "Would you be able to distill the essence of this image into one detailed paragraph of exactly 30 words?", |
| | "Can you unpack this image's details in a succinct paragraph, ensuring it's contained to up to 30 words?", |
| | "Could you elaborate on what this image shows, using no more than 30 words, all within one paragraph?", |
| | "In 30 words or fewer, can you dissect the details of this image, presented in a single paragraph?", |
| | "Could you render a vivid description of this image within the confines of 30 words, all in one paragraph?", |
| | "Please distill the details of this image into a brief yet rich description, not exceeding 30 words, all in one paragraph.", |
| | "Can you encapsulate this image's details in a comprehensive paragraph, without exceeding 30 words?", |
| | "Would you mind providing a detailed explanation of this image, adhering to a 30-word limit, all in one paragraph?", |
| | "Could you convey the intricate details of this image in a brief composition of no more than 30 words, contained in one paragraph?", |
| | "Please craft a detailed depiction of this image, ensuring it's concise with a maximum of 30 words, all within a single paragraph.", |
| | "Can you delineate the specifics of this image in a succinct narrative, capped at 30 words, all in one paragraph?", |
| | ] |
| |
|
| | landmark_prompt = [ |
| | "In the context of: {global_caption} Try to categorize the following image into one of these categories ['indoor', 'outdoor'], use 'others' if it is not a natural image." |
| | ] |
| |
|
| |
|
| | class CaptionImage: |
| | def __init__(self, cfg, logger, device, init_lava=False): |
| | self.cfg = cfg |
| | self.logger = logger |
| | self.device = device |
| |
|
| | |
| | if init_lava: |
| | from osdsynth.processor.wrappers.llava import LLavaWrapper |
| |
|
| | self.llava_processor = LLavaWrapper(cfg, logger, device) |
| | else: |
| | self.llava_processor = None |
| |
|
| | def process_landmark(self, image_bgr): |
| | image_tensor, image_size = self.llava_processor.process_image(image_bgr) |
| |
|
| | global_qs = random.choice(global_qs_list) |
| | global_caption = self.llava_processor.process_vqa(image_tensor, image_size, global_qs, 1024) |
| |
|
| | landmark_qs = random.choice(landmark_prompt) |
| | landmark_qs = landmark_qs.format(global_caption=global_caption) |
| | landmark_caption = self.llava_processor.process_vqa(image_tensor, image_size, landmark_qs, 50) |
| |
|
| | if "indoor" in landmark_caption.lower(): |
| | landmark = "indoor" |
| | elif "outdoor" in landmark_caption.lower(): |
| | landmark = "outdoor" |
| | else: |
| | raise SkipImageException("LLava failed to predict the landmark.") |
| |
|
| | return landmark, global_caption |
| |
|
| | def process_local_caption(self, detections, global_caption="", use_placeholder=True, is_one=False, is_three=False): |
| | n_objects = len(detections) |
| | if not is_three: |
| | if n_objects < 2 and is_one==False: |
| | raise SkipImageException("Ddetected objects less than 2") |
| | if n_objects == 0: |
| | raise SkipImageException("No objects detected finally") |
| | else: |
| | if n_objects < 3: |
| | raise SkipImageException("Ddetected objects less than 3") |
| | if n_objects == 0: |
| | raise SkipImageException("No objects detected finally") |
| |
|
| | for obj_idx in range(n_objects): |
| | if use_placeholder: |
| | |
| | detections[obj_idx]["caption"] = f"<region{obj_idx}: {detections[obj_idx]['class_name']}>" |
| | else: |
| | assert self.llava_processor is not None |
| | detections[obj_idx]["caption"] = f"<region{obj_idx}>" |
| |
|
| | |
| | cropped_image = detections[obj_idx]["image_crop_modified"] |
| | image_tensor, image_size = self.llava_processor.process_image(cropped_image, is_pil_rgb=True) |
| |
|
| | local_qs_template = r""""Can you describe the {class_name} in this close-up within five words? Highlight its color, appearance, style. For example: 'Man in red hat walking', 'Wooden pallet with boxes'.""" |
| | local_qs = local_qs_template.format( |
| | global_caption=global_caption, class_name=detections[obj_idx]["class_name"] |
| | ) |
| | local_caption = self.llava_processor.process_vqa(image_tensor, image_size, local_qs, 50) |
| | detections[obj_idx]["dense_caption"] = local_caption.lower().replace(".", "").strip('"') |
| | return detections |
| |
|
