230 lines
8.6 KiB
Python
230 lines
8.6 KiB
Python
# Copyright 2023 DeepMind Technologies Limited
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
import io
|
|
import glob
|
|
import torch
|
|
import pickle as pkl
|
|
import numpy as np
|
|
import os.path as osp
|
|
import mediapy as media
|
|
from torch.utils.data import Dataset, DataLoader
|
|
|
|
from PIL import Image
|
|
from typing import Mapping, Tuple
|
|
|
|
|
|
def resize_video(video: np.ndarray, output_size: Tuple[int, int]) -> np.ndarray:
|
|
"""Resize a video to output_size."""
|
|
# If you have a GPU, consider replacing this with a GPU-enabled resize op,
|
|
# such as a jitted jax.image.resize. It will make things faster.
|
|
return media.resize_video(video, output_size)
|
|
|
|
|
|
def sample_queries_first(
|
|
target_occluded: np.ndarray,
|
|
target_points: np.ndarray,
|
|
frames: np.ndarray,
|
|
) -> Mapping[str, np.ndarray]:
|
|
"""Package a set of frames and tracks for use in TAPNet evaluations.
|
|
Given a set of frames and tracks with no query points, use the first
|
|
visible point in each track as the query.
|
|
Args:
|
|
target_occluded: Boolean occlusion flag, of shape [n_tracks, n_frames],
|
|
where True indicates occluded.
|
|
target_points: Position, of shape [n_tracks, n_frames, 2], where each point
|
|
is [x,y] scaled between 0 and 1.
|
|
frames: Video tensor, of shape [n_frames, height, width, 3]. Scaled between
|
|
-1 and 1.
|
|
Returns:
|
|
A dict with the keys:
|
|
video: Video tensor of shape [1, n_frames, height, width, 3]
|
|
query_points: Query points of shape [1, n_queries, 3] where
|
|
each point is [t, y, x] scaled to the range [-1, 1]
|
|
target_points: Target points of shape [1, n_queries, n_frames, 2] where
|
|
each point is [x, y] scaled to the range [-1, 1]
|
|
"""
|
|
valid = np.sum(~target_occluded, axis=1) > 0
|
|
target_points = target_points[valid, :]
|
|
target_occluded = target_occluded[valid, :]
|
|
|
|
query_points = []
|
|
for i in range(target_points.shape[0]):
|
|
index = np.where(target_occluded[i] == 0)[0][0]
|
|
x, y = target_points[i, index, 0], target_points[i, index, 1]
|
|
query_points.append(np.array([index, y, x])) # [t, y, x]
|
|
query_points = np.stack(query_points, axis=0)
|
|
|
|
return {
|
|
"video": frames[np.newaxis, ...],
|
|
"query_points": query_points[np.newaxis, ...],
|
|
"target_points": target_points[np.newaxis, ...],
|
|
"occluded": target_occluded[np.newaxis, ...],
|
|
}
|
|
|
|
|
|
def sample_queries_strided(
|
|
target_occluded: np.ndarray,
|
|
target_points: np.ndarray,
|
|
frames: np.ndarray,
|
|
query_stride: int = 5,
|
|
) -> Mapping[str, np.ndarray]:
|
|
"""Package a set of frames and tracks for use in TAPNet evaluations.
|
|
|
|
Given a set of frames and tracks with no query points, sample queries
|
|
strided every query_stride frames, ignoring points that are not visible
|
|
at the selected frames.
|
|
|
|
Args:
|
|
target_occluded: Boolean occlusion flag, of shape [n_tracks, n_frames],
|
|
where True indicates occluded.
|
|
target_points: Position, of shape [n_tracks, n_frames, 2], where each point
|
|
is [x,y] scaled between 0 and 1.
|
|
frames: Video tensor, of shape [n_frames, height, width, 3]. Scaled between
|
|
-1 and 1.
|
|
query_stride: When sampling query points, search for un-occluded points
|
|
every query_stride frames and convert each one into a query.
|
|
|
|
Returns:
|
|
A dict with the keys:
|
|
video: Video tensor of shape [1, n_frames, height, width, 3]. The video
|
|
has floats scaled to the range [-1, 1].
|
|
query_points: Query points of shape [1, n_queries, 3] where
|
|
each point is [t, y, x] scaled to the range [-1, 1].
|
|
target_points: Target points of shape [1, n_queries, n_frames, 2] where
|
|
each point is [x, y] scaled to the range [-1, 1].
|
|
trackgroup: Index of the original track that each query point was
|
|
sampled from. This is useful for visualization.
|
|
"""
|
|
tracks = []
|
|
occs = []
|
|
queries = []
|
|
trackgroups = []
|
|
total = 0
|
|
trackgroup = np.arange(target_occluded.shape[0])
|
|
for i in range(0, target_occluded.shape[1], query_stride):
|
|
mask = target_occluded[:, i] == 0
|
|
query = np.stack(
|
|
[
|
|
i * np.ones(target_occluded.shape[0:1]),
|
|
target_points[:, i, 1],
|
|
target_points[:, i, 0],
|
|
],
|
|
axis=-1,
|
|
)
|
|
queries.append(query[mask])
|
|
tracks.append(target_points[mask])
|
|
occs.append(target_occluded[mask])
|
|
trackgroups.append(trackgroup[mask])
|
|
total += np.array(np.sum(target_occluded[:, i] == 0))
|
|
|
|
return {
|
|
"video": frames[np.newaxis, ...],
|
|
"query_points": np.concatenate(queries, axis=0)[np.newaxis, ...],
|
|
"target_points": np.concatenate(tracks, axis=0)[np.newaxis, ...],
|
|
"occluded": np.concatenate(occs, axis=0)[np.newaxis, ...],
|
|
"trackgroup": np.concatenate(trackgroups, axis=0)[np.newaxis, ...],
|
|
}
|
|
|
|
|
|
class TapVid(Dataset):
|
|
def __init__(
|
|
self,
|
|
data_root,
|
|
split="davis",
|
|
query_mode="first",
|
|
resize_to_256=True
|
|
):
|
|
self.split = split
|
|
self.resize_to_256 = resize_to_256
|
|
self.query_mode = query_mode
|
|
if self.split == "kinetics":
|
|
all_paths = glob.glob(osp.join(data_root, "*_of_0010.pkl"))
|
|
points_dataset = []
|
|
for pickle_path in all_paths:
|
|
with open(pickle_path, "rb") as f:
|
|
data = pkl.load(f)
|
|
points_dataset = points_dataset + data
|
|
self.points_dataset = points_dataset
|
|
else:
|
|
with open(data_root, "rb") as f:
|
|
self.points_dataset = pkl.load(f)
|
|
if self.split == "davis":
|
|
self.video_names = list(self.points_dataset.keys())
|
|
print("found %d unique videos in %s" % (len(self.points_dataset), data_root))
|
|
|
|
def __getitem__(self, index):
|
|
if self.split == "davis":
|
|
video_name = self.video_names[index]
|
|
else:
|
|
video_name = index
|
|
video = self.points_dataset[video_name]
|
|
frames = video["video"]
|
|
|
|
if isinstance(frames[0], bytes):
|
|
# TAP-Vid is stored and JPEG bytes rather than `np.ndarray`s.
|
|
def decode(frame):
|
|
byteio = io.BytesIO(frame)
|
|
img = Image.open(byteio)
|
|
return np.array(img)
|
|
|
|
frames = np.array([decode(frame) for frame in frames])
|
|
|
|
target_points = self.points_dataset[video_name]["points"]
|
|
if self.resize_to_256:
|
|
frames = resize_video(frames, [256, 256])
|
|
target_points *= np.array([256, 256])
|
|
else:
|
|
target_points *= np.array([frames.shape[2], frames.shape[1]])
|
|
|
|
target_occ = self.points_dataset[video_name]["occluded"]
|
|
if self.query_mode == "first":
|
|
converted = sample_queries_first(target_occ, target_points, frames)
|
|
else:
|
|
converted = sample_queries_strided(target_occ, target_points, frames)
|
|
assert converted["target_points"].shape[1] == converted["query_points"].shape[1]
|
|
|
|
trajs = torch.from_numpy(converted["target_points"])[0].permute(1, 0, 2).float() # T, N, D
|
|
|
|
rgbs = torch.from_numpy(frames).permute(0, 3, 1, 2).float() / 255.
|
|
visibles = torch.logical_not(torch.from_numpy(converted["occluded"]))[0].permute(1, 0) # T, N
|
|
query_points = torch.from_numpy(converted["query_points"])[0].float() # T, N
|
|
tracks = torch.cat([trajs, visibles[..., None]], dim=-1)
|
|
|
|
data = {
|
|
"video": rgbs,
|
|
"query_points": query_points,
|
|
"tracks": tracks
|
|
}
|
|
|
|
return data
|
|
|
|
def __len__(self):
|
|
return len(self.points_dataset)
|
|
|
|
|
|
def create_point_tracking_dataset(args):
|
|
data_root = osp.join(args.data_root, f"tapvid_{args.split}")
|
|
if args.split in ["davis", "rgb_stacking"]:
|
|
data_root = osp.join(data_root, f"tapvid_{args.split}.pkl")
|
|
dataset = TapVid(data_root, args.split, args.query_mode)
|
|
dataloader = DataLoader(
|
|
dataset,
|
|
batch_size=args.batch_size,
|
|
pin_memory=True,
|
|
shuffle=False,
|
|
num_workers=0,
|
|
drop_last=False,
|
|
)
|
|
return dataloader |