Qnancy's picture
Add files using upload-large-folder tool
084d74a verified
raw
history blame
12.4 kB
import os
import sys
from typing import Any, Dict, Generator, ItemsView, List, Tuple
import cv2
import numpy as np
import torch
from PIL import Image
GSA_PATH = "/mnt/prev_nas/qhy_1/GenSpace/osdsynth/external/Grounded-Segment-Anything"
sys.path.append(GSA_PATH)
from segment_anything.segment_anything import SamAutomaticMaskGenerator, SamPredictor, sam_hq_model_registry, sam_model_registry
# Segment-Anything checkpoint
SAM_ENCODER_VERSION = "vit_h"
SAM_CHECKPOINT_PATH = os.path.join(GSA_PATH, "./sam_vit_h_4b8939.pth")
# Segment-Anything checkpoint
SAM_HQ_ENCODER_VERSION = "vit_h"
SAM_HQ_CHECKPOINT_PATH = os.path.join(GSA_PATH, "./sam_hq_vit_h.pth")
# Prompting SAM with detected boxes
def get_sam_segmentation_from_xyxy(sam_predictor: SamPredictor, image: np.ndarray, xyxy: np.ndarray) -> np.ndarray:
sam_predictor.set_image(image)
result_masks = []
for box in xyxy:
masks, scores, logits = sam_predictor.predict(box=box, multimask_output=True)
index = np.argmax(scores)
result_masks.append(masks[index])
return np.array(result_masks)
def get_sam_predictor(variant: str, device: str | int) -> SamPredictor:
if variant == "sam":
sam = sam_model_registry[SAM_ENCODER_VERSION](checkpoint=SAM_CHECKPOINT_PATH)
sam.to(device)
sam_predictor = SamPredictor(sam)
return sam_predictor
if variant == "sam-hq":
print("Using SAM-HQ")
sam = sam_hq_model_registry[SAM_HQ_ENCODER_VERSION](checkpoint=SAM_HQ_CHECKPOINT_PATH)
sam.to(device)
sam_predictor = SamPredictor(sam)
return sam_predictor
else:
raise NotImplementedError
def get_sam_mask_generator(variant: str, device: str | int) -> SamAutomaticMaskGenerator:
if variant == "sam":
sam = sam_model_registry[SAM_ENCODER_VERSION](checkpoint=SAM_CHECKPOINT_PATH)
sam.to(device)
mask_generator = SamAutomaticMaskGenerator(
model=sam,
points_per_side=12,
points_per_batch=144,
pred_iou_thresh=0.88,
stability_score_thresh=0.95,
crop_n_layers=0,
min_mask_region_area=100,
)
return mask_generator
elif variant == "fastsam":
raise NotImplementedError
else:
raise NotImplementedError
def convert_detections_to_list(detections_dict, classes):
detection_list = []
for i in range(len(detections_dict["xyxy"])):
detection = {
"class_name": classes[detections_dict["class_id"][i]], # Lookup class name using class_id
"xyxy": detections_dict["xyxy"][i], # Assuming detections.xyxy is a numpy array
"confidence": detections_dict["confidence"][i].item(), # Convert numpy scalar to Python scalar
"class_id": detections_dict["class_id"][i].item(),
"box_area": detections_dict["box_area"][i].item(),
"mask": detections_dict["mask"][i],
"subtracted_mask": detections_dict["subtracted_mask"][i],
"rle": detections_dict["rle"][i],
"area": detections_dict["area"][i],
}
detection_list.append(detection)
return detection_list
def convert_detections_to_dict(detections, classes, image_crops=None, image_feats=None, text_feats=None):
# Convert the detections to a dict. The elements are in np.array
results = {
"xyxy": detections.xyxy,
"confidence": detections.confidence,
"class_id": detections.class_id,
"box_area": detections.box_area,
"mask": detections.mask,
"area": detections.area,
"classes": classes,
}
return results
def mask_subtract_contained(xyxy: np.ndarray, mask: np.ndarray, th1=0.8, th2=0.7):
"""Compute the containing relationship between all pair of bounding boxes. For each mask, subtract the mask of
bounding boxes that are contained by it.
Args:
xyxy: (N, 4), in (x1, y1, x2, y2) format
mask: (N, H, W), binary mask
th1: float, threshold for computing intersection over box1
th2: float, threshold for computing intersection over box2
Returns:
mask_sub: (N, H, W), binary mask
"""
N = xyxy.shape[0] # number of boxes
# Get areas of each xyxy
areas = (xyxy[:, 2] - xyxy[:, 0]) * (xyxy[:, 3] - xyxy[:, 1]) # (N,)
# Compute intersection boxes
lt = np.maximum(xyxy[:, None, :2], xyxy[None, :, :2]) # left-top points (N, N, 2)
rb = np.minimum(xyxy[:, None, 2:], xyxy[None, :, 2:]) # right-bottom points (N, N, 2)
inter = (rb - lt).clip(min=0) # intersection sizes (dx, dy), if no overlap, clamp to zero (N, N, 2)
# Compute areas of intersection boxes
inter_areas = inter[:, :, 0] * inter[:, :, 1] # (N, N)
inter_over_box1 = inter_areas / areas[:, None] # (N, N)
# inter_over_box2 = inter_areas / areas[None, :] # (N, N)
inter_over_box2 = inter_over_box1.T # (N, N)
# if the intersection area is smaller than th2 of the area of box1,
# and the intersection area is larger than th1 of the area of box2,
# then box2 is considered contained by box1
contained = (inter_over_box1 < th2) & (inter_over_box2 > th1) # (N, N)
contained_idx = contained.nonzero() # (num_contained, 2)
mask_sub = mask.copy() # (N, H, W)
# mask_sub[contained_idx[0]] = mask_sub[contained_idx[0]] & (~mask_sub[contained_idx[1]])
for i in range(len(contained_idx[0])):
mask_sub[contained_idx[0][i]] = mask_sub[contained_idx[0][i]] & (~mask_sub[contained_idx[1][i]])
return mask_sub, contained
def filter_detections(cfg, detections_dict: dict, image: np.ndarray):
# If no detection at all
if len(detections_dict["xyxy"]) == 0:
return detections_dict
# Filter out the objects based on various criteria
idx_to_keep = []
for obj_idx in range(len(detections_dict["xyxy"])):
class_name = detections_dict["classes"][detections_dict["class_id"][obj_idx]]
# Skip masks that are too small
if detections_dict["mask"][obj_idx].sum() < max(cfg.mask_area_threshold, 10):
print(f"Skipping {class_name} mask with too few points")
continue
# Skip the BG classes
if cfg.skip_bg and class_name in cfg.bg_classes:
print(f"Skipping {class_name} as it is a background class")
continue
# Skip the non-background boxes that are too large
if class_name not in cfg.bg_classes:
x1, y1, x2, y2 = detections_dict["xyxy"][obj_idx]
bbox_area = (x2 - x1) * (y2 - y1)
image_area = image.shape[0] * image.shape[1]
if bbox_area > cfg.max_bbox_area_ratio * image_area:
print(f"Skipping {class_name} with area {bbox_area} > {cfg.max_bbox_area_ratio} * {image_area}")
continue
# Skip masks with low confidence
if detections_dict["confidence"][obj_idx] < cfg.mask_conf_threshold:
print(
f"Skipping {class_name} with confidence {detections_dict['confidence'][obj_idx]} < {cfg.mask_conf_threshold}"
)
continue
idx_to_keep.append(obj_idx)
for k in detections_dict.keys():
if isinstance(detections_dict[k], str) or k == "classes": # Captions
continue
elif isinstance(detections_dict[k], list):
detections_dict[k] = [detections_dict[k][i] for i in idx_to_keep]
elif isinstance(detections_dict[k], np.ndarray):
detections_dict[k] = detections_dict[k][idx_to_keep]
else:
raise NotImplementedError(f"Unhandled type {type(detections_dict[k])}")
return detections_dict
def sort_detections_by_area(detections_dict):
# Sort the detections by area, use negative to sort from large to small
sorted_indices = np.argsort(-detections_dict["area"])
for key in detections_dict.keys():
if isinstance(detections_dict[key], np.ndarray): # Check to ensure it's an array
detections_dict[key] = detections_dict[key][sorted_indices]
return detections_dict
def post_process_mask(detections_dict):
sam_masks = torch.tensor(detections_dict["subtracted_mask"])
uncompressed_mask_rles = mask_to_rle_pytorch(sam_masks)
rle_masks_list = [coco_encode_rle(uncompressed_mask_rles[i]) for i in range(len(uncompressed_mask_rles))]
detections_dict["rle"] = rle_masks_list
return detections_dict
def crop_image_and_mask(image: Image, mask: np.ndarray, x1: int, y1: int, x2: int, y2: int, padding: int = 0):
"""Crop the image and mask with some padding.
I made a single function that crops both the image and the mask at the same time because I was getting shape
mismatches when I cropped them separately.This way I can check that they are the same shape.
"""
image = np.array(image)
# Verify initial dimensions
if image.shape[:2] != mask.shape:
print(f"Initial shape mismatch: Image shape {image.shape} != Mask shape {mask.shape}")
return None, None
# Define the cropping coordinates
x1 = max(0, x1 - padding)
y1 = max(0, y1 - padding)
x2 = min(image.shape[1], x2 + padding)
y2 = min(image.shape[0], y2 + padding)
# round the coordinates to integers
x1, y1, x2, y2 = round(x1), round(y1), round(x2), round(y2)
# Crop the image and the mask
image_crop = image[y1:y2, x1:x2]
mask_crop = mask[y1:y2, x1:x2]
# Verify cropped dimensions
if image_crop.shape[:2] != mask_crop.shape:
print(
"Cropped shape mismatch: Image crop shape {} != Mask crop shape {}".format(
image_crop.shape, mask_crop.shape
)
)
return None, None
# convert the image back to a pil image
image_crop = Image.fromarray(image_crop)
return image_crop, mask_crop
def crop_detections_with_xyxy(cfg, image, detections_list):
for idx, detection in enumerate(detections_list):
x1, y1, x2, y2 = detection["xyxy"]
image_crop, mask_crop = crop_image_and_mask(image, detection["mask"], x1, y1, x2, y2, padding=10)
if cfg.masking_option == "blackout":
image_crop_modified = blackout_nonmasked_area(image_crop, mask_crop)
elif cfg.masking_option == "red_outline":
image_crop_modified = draw_red_outline(image_crop, mask_crop)
else:
image_crop_modified = image_crop # No modification
detections_list[idx]["image_crop"] = image_crop
detections_list[idx]["mask_crop"] = mask_crop
detections_list[idx]["image_crop_modified"] = image_crop_modified
return detections_list
def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
"""
Encodes masks to an uncompressed RLE, in the format expected by
pycoco tools.
"""
# Put in fortran order and flatten h,w
b, h, w = tensor.shape
tensor = tensor.permute(0, 2, 1).flatten(1)
# Compute change indices
diff = tensor[:, 1:] ^ tensor[:, :-1]
change_indices = diff.nonzero()
# Encode run length
out = []
for i in range(b):
cur_idxs = change_indices[change_indices[:, 0] == i, 1]
cur_idxs = torch.cat(
[
torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
cur_idxs + 1,
torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
]
)
btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
counts = [] if tensor[i, 0] == 0 else [0]
counts.extend(btw_idxs.detach().cpu().tolist())
out.append({"size": [h, w], "counts": counts})
return out
def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
"""Compute a binary mask from an uncompressed RLE."""
h, w = rle["size"]
mask = np.empty(h * w, dtype=bool)
idx = 0
parity = False
for count in rle["counts"]:
mask[idx : idx + count] = parity
idx += count
parity ^= True
mask = mask.reshape(w, h)
return mask.transpose() # Put in C order
def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
from pycocotools import mask as mask_utils # type: ignore
h, w = uncompressed_rle["size"]
rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
rle["counts"] = rle["counts"].decode("utf-8") # Necessary to serialize with json
return rle