|
| 1 | +# Copyright (C) CVAT.ai Corporation |
| 2 | +# |
| 3 | +# SPDX-License-Identifier: MIT |
| 4 | + |
| 5 | +import collections |
| 6 | +import dataclasses |
| 7 | +from typing import Optional, TypedDict |
| 8 | + |
| 9 | +import cv2 |
| 10 | +import cvat_sdk.auto_annotation as cvataa |
| 11 | +import numpy as np |
| 12 | +import PIL.Image |
| 13 | +import torch |
| 14 | +import torchvision.transforms |
| 15 | +from cvat_sdk.masks import decode_mask, encode_mask |
| 16 | +from sam2.sam2_video_predictor import SAM2VideoPredictor |
| 17 | +from sam2.utils.misc import fill_holes_in_mask_scores |
| 18 | + |
| 19 | + |
| 20 | +@dataclasses.dataclass(frozen=True, kw_only=True) |
| 21 | +class _PreprocessedImage: |
| 22 | + original_width: int |
| 23 | + original_height: int |
| 24 | + vision_feats: list[torch.Tensor] |
| 25 | + vision_pos_embeds: list[torch.Tensor] |
| 26 | + feat_sizes: list[tuple[int, int]] |
| 27 | + |
| 28 | + |
| 29 | +class _PredictorOutputs(TypedDict): |
| 30 | + # We always keep 1 cond_frame_outputs and up to num_maskmem non_cond_frame_outputs. |
| 31 | + |
| 32 | + cond_frame_outputs: dict[int, dict] |
| 33 | + # We make this an OrderedDict to make popping old elements easier. |
| 34 | + non_cond_frame_outputs: collections.OrderedDict[int, dict] |
| 35 | + |
| 36 | + |
| 37 | +@dataclasses.dataclass(kw_only=True) |
| 38 | +class _TrackingState: |
| 39 | + frame_idx: int |
| 40 | + predictor_outputs: _PredictorOutputs |
| 41 | + |
| 42 | + |
| 43 | +class _Sam2Tracker: |
| 44 | + def __init__(self, model_id: str, device: str = "cpu", **kwargs) -> None: |
| 45 | + self._device = torch.device(device) |
| 46 | + |
| 47 | + if self._device.type == "cuda": |
| 48 | + torch.set_autocast_enabled(True) |
| 49 | + torch.set_autocast_gpu_dtype(torch.bfloat16) |
| 50 | + if torch.cuda.get_device_properties(self._device).major >= 8: |
| 51 | + torch.backends.cuda.matmul.allow_tf32 = True |
| 52 | + torch.backends.cudnn.allow_tf32 = True |
| 53 | + |
| 54 | + self._predictor = SAM2VideoPredictor.from_pretrained( |
| 55 | + model_id, device=self._device, **kwargs |
| 56 | + ) |
| 57 | + self._transform = torchvision.transforms.Compose( |
| 58 | + [ |
| 59 | + torchvision.transforms.Resize( |
| 60 | + (self._predictor.image_size, self._predictor.image_size) |
| 61 | + ), |
| 62 | + torchvision.transforms.ToTensor(), |
| 63 | + torchvision.transforms.Normalize( |
| 64 | + # see load_video_frames in the SAM2 source |
| 65 | + mean=(0.485, 0.456, 0.406), |
| 66 | + std=(0.229, 0.224, 0.225), |
| 67 | + ), |
| 68 | + ] |
| 69 | + ) |
| 70 | + |
| 71 | + spec = cvataa.TrackingFunctionSpec(supported_shape_types=["mask", "polygon"]) |
| 72 | + |
| 73 | + @torch.inference_mode() |
| 74 | + def preprocess_image( |
| 75 | + self, context: cvataa.TrackingFunctionContext, image: PIL.Image.Image |
| 76 | + ) -> _PreprocessedImage: |
| 77 | + image = image.convert("RGB") |
| 78 | + |
| 79 | + image_tensor = self._transform(image).unsqueeze(0).to(device=self._device) |
| 80 | + backbone_out = self._predictor.forward_image(image_tensor) |
| 81 | + vision_feats = backbone_out["backbone_fpn"][-self._predictor.num_feature_levels :] |
| 82 | + vision_pos_embeds = backbone_out["vision_pos_enc"][-self._predictor.num_feature_levels :] |
| 83 | + |
| 84 | + return _PreprocessedImage( |
| 85 | + original_width=image.width, |
| 86 | + original_height=image.height, |
| 87 | + vision_feats=[x.flatten(2).permute(2, 0, 1) for x in vision_feats], |
| 88 | + vision_pos_embeds=[x.flatten(2).permute(2, 0, 1) for x in vision_pos_embeds], |
| 89 | + feat_sizes=[(x.shape[-2], x.shape[-1]) for x in vision_pos_embeds], |
| 90 | + ) |
| 91 | + |
| 92 | + def _call_predictor(self, *, pp_image: _PreprocessedImage, frame_idx: int, **kwargs) -> dict: |
| 93 | + out = self._predictor.track_step( |
| 94 | + current_vision_feats=pp_image.vision_feats, |
| 95 | + current_vision_pos_embeds=pp_image.vision_pos_embeds, |
| 96 | + feat_sizes=pp_image.feat_sizes, |
| 97 | + point_inputs=None, |
| 98 | + frame_idx=frame_idx, |
| 99 | + num_frames=frame_idx + 1, |
| 100 | + **kwargs, |
| 101 | + ) |
| 102 | + |
| 103 | + return { |
| 104 | + "maskmem_features": out["maskmem_features"], |
| 105 | + "maskmem_pos_enc": out["maskmem_pos_enc"][-1:], |
| 106 | + "pred_masks": fill_holes_in_mask_scores( |
| 107 | + out["pred_masks"], self._predictor.fill_hole_area |
| 108 | + ), |
| 109 | + "obj_ptr": out["obj_ptr"], |
| 110 | + } |
| 111 | + |
| 112 | + def _shape_to_mask( |
| 113 | + self, pp_image: _PreprocessedImage, shape: cvataa.TrackableShape |
| 114 | + ) -> np.ndarray: |
| 115 | + if shape.type == "mask": |
| 116 | + return decode_mask( |
| 117 | + shape.points, |
| 118 | + image_width=pp_image.original_width, |
| 119 | + image_height=pp_image.original_height, |
| 120 | + ) |
| 121 | + |
| 122 | + if shape.type == "polygon": |
| 123 | + mask = np.zeros((pp_image.original_height, pp_image.original_width), dtype=np.uint8) |
| 124 | + points_array = np.array(shape.points, dtype=np.int32).reshape((-1, 2)) |
| 125 | + cv2.fillPoly(mask, [points_array], 1) |
| 126 | + return mask |
| 127 | + |
| 128 | + assert False, f"unexpected shape type {shape.type!r}" |
| 129 | + |
| 130 | + @torch.inference_mode() |
| 131 | + def init_tracking_state( |
| 132 | + self, |
| 133 | + context: cvataa.TrackingFunctionShapeContext, |
| 134 | + pp_image: _PreprocessedImage, |
| 135 | + shape: cvataa.TrackableShape, |
| 136 | + ) -> _TrackingState: |
| 137 | + mask = torch.from_numpy(self._shape_to_mask(pp_image, shape)) |
| 138 | + |
| 139 | + resized_mask = torch.nn.functional.interpolate( |
| 140 | + mask.float()[None, None], # add batch and channel dimensions |
| 141 | + (self._predictor.image_size, self._predictor.image_size), |
| 142 | + mode="bilinear", |
| 143 | + align_corners=False, |
| 144 | + ) |
| 145 | + resized_mask = (resized_mask >= 0.5).float().to(device=self._device) |
| 146 | + |
| 147 | + current_out = self._call_predictor( |
| 148 | + pp_image=pp_image, |
| 149 | + frame_idx=0, |
| 150 | + is_init_cond_frame=True, |
| 151 | + mask_inputs=resized_mask, |
| 152 | + output_dict={}, |
| 153 | + ) |
| 154 | + |
| 155 | + return _TrackingState( |
| 156 | + frame_idx=0, |
| 157 | + predictor_outputs={ |
| 158 | + "cond_frame_outputs": {0: current_out}, |
| 159 | + "non_cond_frame_outputs": collections.OrderedDict(), |
| 160 | + }, |
| 161 | + ) |
| 162 | + |
| 163 | + def _mask_to_shape( |
| 164 | + self, context: cvataa.TrackingFunctionShapeContext, mask: torch.Tensor |
| 165 | + ) -> Optional[cvataa.TrackableShape]: |
| 166 | + if context.original_shape_type == "mask": |
| 167 | + return cvataa.TrackableShape(type="mask", points=encode_mask(mask)) |
| 168 | + |
| 169 | + if context.original_shape_type == "polygon": |
| 170 | + mask_np = np.asarray(mask, dtype=np.uint8) |
| 171 | + contours, _ = cv2.findContours(mask_np, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) |
| 172 | + if not contours: |
| 173 | + return None |
| 174 | + |
| 175 | + largest_contour = max(contours, key=cv2.contourArea) |
| 176 | + approx_contour = cv2.approxPolyDP(largest_contour, epsilon=1.0, closed=True) |
| 177 | + if approx_contour.shape[0] < 3: |
| 178 | + return None |
| 179 | + |
| 180 | + return cvataa.TrackableShape(type="polygon", points=approx_contour.flatten().tolist()) |
| 181 | + |
| 182 | + assert False, f"unexpected shape type {context.original_shape_type!r}" |
| 183 | + |
| 184 | + @torch.inference_mode() |
| 185 | + def track( |
| 186 | + self, |
| 187 | + context: cvataa.TrackingFunctionShapeContext, |
| 188 | + pp_image: _PreprocessedImage, |
| 189 | + state: _TrackingState, |
| 190 | + ) -> Optional[cvataa.TrackableShape]: |
| 191 | + state.frame_idx += 1 |
| 192 | + |
| 193 | + current_out = self._call_predictor( |
| 194 | + pp_image=pp_image, |
| 195 | + frame_idx=state.frame_idx, |
| 196 | + is_init_cond_frame=False, |
| 197 | + mask_inputs=None, |
| 198 | + output_dict=state.predictor_outputs, |
| 199 | + ) |
| 200 | + |
| 201 | + non_cond_frame_outputs = state.predictor_outputs["non_cond_frame_outputs"] |
| 202 | + non_cond_frame_outputs[state.frame_idx] = current_out |
| 203 | + |
| 204 | + # discard old outputs as the predictor uses up to num_maskmem elements |
| 205 | + while len(non_cond_frame_outputs) > self._predictor.num_maskmem: |
| 206 | + non_cond_frame_outputs.popitem(last=False) |
| 207 | + |
| 208 | + output_mask = ( |
| 209 | + torch.nn.functional.interpolate( |
| 210 | + current_out["pred_masks"], |
| 211 | + size=(pp_image.original_height, pp_image.original_width), |
| 212 | + align_corners=False, |
| 213 | + mode="bilinear", |
| 214 | + antialias=True, |
| 215 | + )[0, 0] |
| 216 | + > 0 |
| 217 | + ) |
| 218 | + |
| 219 | + if output_mask.any(): |
| 220 | + return self._mask_to_shape(context, output_mask.cpu()) |
| 221 | + else: |
| 222 | + return None |
| 223 | + |
| 224 | + |
| 225 | +create = _Sam2Tracker |
![cvat-bot[bot]](https://avatars.githubusercontent.com/in/406724?v=4&size=40){kind=avatar}
0 commit comments