Spaces:
Running
on
Zero
Running
on
Zero
File size: 10,530 Bytes
9b33fca |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 |
"""SHIFT result writer."""
from __future__ import annotations
import io
import itertools
import json
import os
from collections import defaultdict
import numpy as np
from PIL import Image
from vis4d.common.array import array_to_numpy
from vis4d.common.imports import SCALABEL_AVAILABLE
from vis4d.common.typing import (
ArrayLike,
GenericFunc,
MetricLogs,
NDArrayNumber,
)
from vis4d.data.datasets.shift import shift_det_map
from vis4d.data.io import DataBackend, ZipBackend
from vis4d.eval.base import Evaluator
if SCALABEL_AVAILABLE:
from scalabel.label.transforms import mask_to_rle, xyxy_to_box2d
from scalabel.label.typing import Dataset, Frame, Label
else:
raise ImportError("scalabel is not installed.")
class SHIFTMultitaskWriter(Evaluator):
"""SHIFT result writer for online evaluation."""
inverse_cat_map = {v: k for k, v in shift_det_map.items()}
def __init__(
self,
output_dir: str,
submission_file: str = "submission.zip",
) -> None:
"""Creates a new writer.
Args:
output_dir (str): Output directory.
submission_file (str): Submission file name. Defaults to
"submission.zip".
"""
super().__init__()
assert submission_file.endswith(
".zip"
), "Submission file must be a zip file."
self.backend: DataBackend = ZipBackend()
self.output_path = os.path.join(output_dir, submission_file)
self.frames_det_2d: list[Frame] = []
self.frames_det_3d: list[Frame] = []
self.sample_counts: defaultdict[str, int] = defaultdict(int)
def _write_sem_mask(
self, sem_mask: NDArrayNumber, sample_name: str, video_name: str
) -> None:
"""Write semantic mask.
Args:
sem_mask (NDArrayNumber): Predicted semantic mask, shape (H, W).
sample_name (str): Sample name.
video_name (str): Video name.
"""
image = Image.fromarray(sem_mask.astype("uint8"), mode="L")
image_bytes = io.BytesIO()
image.save(image_bytes, format="PNG")
self.backend.set(
f"{self.output_path}/semseg/{video_name}/{sample_name}",
image_bytes.getvalue(),
mode="w",
)
def _write_depth(
self, depth_map: NDArrayNumber, sample_name: str, video_name: str
) -> None:
"""Write depth map.
Args:
depth_map (NDArrayNumber): Predicted depth map, shape (H, W).
sample_name (str): Sample name.
video_name (str): Video name.
"""
depth_map = np.clip(depth_map / 80.0 * 255.0, 0, 255)
image = Image.fromarray(depth_map.astype("uint8"), mode="L")
image_bytes = io.BytesIO()
image.save(image_bytes, format="PNG")
self.backend.set(
f"{self.output_path}/depth/{video_name}/{sample_name}",
image_bytes.getvalue(),
mode="w",
)
def _write_flow(
self, flow: NDArrayNumber, sample_name: str, video_name: str
) -> None:
"""Write semantic mask.
Args:
flow (NDArrayNumber): Predicted optical flow, shape (H, W, 2).
sample_name (str): Sample name.
video_name (str): Video name.
"""
raise NotImplementedError
def process_batch(
self,
frame_ids: list[int],
sample_names: list[str],
sequence_names: list[str],
pred_sem_mask: list[ArrayLike] | None = None,
pred_depth: list[ArrayLike] | None = None,
pred_flow: list[ArrayLike] | None = None,
pred_boxes2d: list[ArrayLike] | None = None,
pred_boxes2d_classes: list[ArrayLike] | None = None,
pred_boxes2d_scores: list[ArrayLike] | None = None,
pred_boxes2d_track_ids: list[ArrayLike] | None = None,
pred_instance_masks: list[ArrayLike] | None = None,
) -> None:
"""Process SHIFT results.
You can omit some of the predictions if they are not used.
Args:
frame_ids (list[int]): Frame IDs.
sample_names (list[str]): Sample names.
sequence_names (list[str]): Sequence names.
pred_sem_mask (list[ArrayLike], optional): Predicted semantic
masks, each in shape (C, H, W) or (H, W). Defaults to None.
pred_depth (list[ArrayLike], optional): Predicted depth maps,
each in shape (H, W), with meter unit. Defaults to None.
pred_flow (list[ArrayLike], optional): Predicted optical flows,
each in shape (H, W, 2). Defaults to None.
pred_boxes2d (list[ArrayLike], optional): Predicted 2D boxes,
each in shape (N, 4). Defaults to None.
pred_boxes2d_classes (list[ArrayLike], optional): Predicted
2D box classes, each in shape (N,). Defaults to None.
pred_boxes2d_scores (list[ArrayLike], optional): Predicted
2D box scores, each in shape (N,). Defaults to None.
pred_boxes2d_track_ids (list[ArrayLike], optional): Predicted
2D box track IDs, each in shape (N,). Defaults to None.
pred_instance_masks (list[ArrayLike], optional): Predicted
instance masks, each in shape (N, H, W). Defaults to None.
"""
for i, (frame_id, sample_name, sequence_name) in enumerate(
zip(frame_ids, sample_names, sequence_names)
):
if pred_sem_mask is not None:
sem_mask_ = array_to_numpy(
pred_sem_mask[i],
n_dims=None,
dtype=np.float32,
)
if len(sem_mask_.shape) == 3:
sem_mask = sem_mask_.argmax(axis=0)
else:
sem_mask = sem_mask_.astype(np.uint8)
semseg_filename = sample_name.replace(".jpg", ".png").replace(
"img", "semseg"
)
self._write_sem_mask(sem_mask, semseg_filename, sequence_name)
self.sample_counts["semseg"] += 1
if pred_depth is not None:
depth = array_to_numpy(
pred_depth[i], n_dims=None, dtype=np.float32
)
depth_filename = sample_name.replace(".jpg", ".png").replace(
"img", "depth"
)
self._write_depth(depth, depth_filename, sequence_name)
self.sample_counts["depth"] += 1
if pred_flow is not None:
flow = array_to_numpy(
pred_flow[i], n_dims=None, dtype=np.float32
)
self._write_flow(flow, sample_name, sequence_name)
self.sample_counts["flow"] += 1
if (
pred_boxes2d is not None
and pred_boxes2d_classes is not None
and pred_boxes2d_scores is not None
):
labels = []
if pred_instance_masks:
masks = array_to_numpy(
pred_instance_masks[i], n_dims=None, dtype=np.float32
)
if pred_boxes2d_track_ids:
track_ids = array_to_numpy(
pred_boxes2d_track_ids[i],
n_dims=None,
dtype=np.int64,
)
for box, score, class_id in zip(
pred_boxes2d[i],
pred_boxes2d_scores[i],
pred_boxes2d_classes[i],
):
box2d = xyxy_to_box2d(*box.tolist())
if pred_instance_masks:
rle = mask_to_rle(
(masks[class_id] > 0.0).astype(np.uint8)
)
else:
rle = None
if pred_boxes2d_track_ids:
track_id = str(int(track_ids[0]))
else:
track_id = None
label = Label(
box2d=box2d,
category=(
self.inverse_cat_map[int(class_id)]
if self.inverse_cat_map != {}
else str(class_id)
),
score=float(score),
rle=rle,
id=track_id,
)
labels.append(label)
frame = Frame(
name=sample_name,
videoName=sequence_name,
frameIndex=frame_id,
labels=labels,
)
self.frames_det_2d.append(frame)
self.sample_counts["det_2d"] += 1
def gather(self, gather_func: GenericFunc) -> None: # pragma: no cover
"""Gather variables in case of distributed setting (if needed).
Args:
gather_func (Callable[[Any], Any]): Gather function.
"""
all_preds = gather_func(self.frames_det_2d)
if all_preds is not None:
self.frames_det_2d = list(itertools.chain(*all_preds))
def evaluate(self, metric: str) -> tuple[MetricLogs, str]:
"""No evaluation locally."""
return {}, "No evaluation locally."
def save(self, metric: str, output_dir: str) -> None:
"""Save scalabel output to zip file.
Raises:
ValueError: If the number of samples in each category is not the
same.
"""
# Check if the sample counts are correct
equal_size = True
for key in self.sample_counts:
if self.sample_counts[key] != len(self.frames_det_2d):
equal_size = False
break
if not equal_size:
raise ValueError(
"The number of samples in each category is not the same."
)
# Save the 2D detection results
if len(self.frames_det_2d) > 0:
ds = Dataset(frames=self.frames_det_2d, groups=None, config=None)
ds_bytes = json.dumps(ds.dict()).encode("utf-8")
self.backend.set(
f"{self.output_path}/det_2d.json", ds_bytes, mode="w"
)
self.backend.close()
print(f"Saved the submission file at {self.output_path}.")
|