a
hR @ st d dl Z d dlZd dlmZmZ d dlmZ d dlmZ d dlm Z m
Z
m Z m Z d dl
Zd dlZd dlmZ ddlmZmZ d d
lmZ d d lmZmZ d d lmZ e ejeje ej e ej f Z e ejeje ej f Z d
Z G dd deeZ!G dd de!Z"G dd de!Z#G dd de!Z$G dd de!Z%G dd de!Z&e'ej ddd Z(e'e ej ej f dd d Z)dS ) N)ABCabstractmethod)globPath)AnyCallableOptionalUnion)Image )
decode_png read_file )default_loader) _read_pfmverify_str_arg)
VisionDataset) KittiFlowSintelFlyingThings3D FlyingChairsHD1Kc s e Zd ZdZdefeeef ee e ege
f dd fdd
Z eee j e
jf dddZeed d
d Zeeeef d d
dZedddZee
jjjdddZ ZS ) FlowDatasetFN)root
transformsloaderreturnc s* t j|d || _g | _g | _|| _d S )N)r )super__init__r
_flow_list _image_list_loader)selfr r r __class__ P/var/www/auris/lib/python3.9/site-packages/torchvision/datasets/_optical_flow.pyr $ s
zFlowDataset.__init__ file_namer c C s
| |S N)r" r# r) r& r& r' _read_img2 s zFlowDataset._read_img)r) c C s d S r* r& r+ r& r& r'
_read_flow5 s zFlowDataset._read_flowindexr c C s | | j| d }| | j| d }| jrT| | j| }| jrN|\}}q\d }nd }}| jd ur~| ||||\}}}}| js|d ur||||fS |||fS d S )Nr r )r, r! r r- _has_builtin_flow_maskr )r# r/ img1img2flowvalid_flow_maskr& r& r' __getitem__: s
zFlowDataset.__getitem__)r c C s
t | jS r* )lenr! )r# r& r& r' __len__Q s zFlowDataset.__len__)vr c C s t jj| g| S r* )torchutilsdata
ConcatDataset)r# r8 r& r& r' __rmul__T s zFlowDataset.__rmul__) __name__
__module__ __qualname__r0 r r
strr r r r r r r9 ZTensorr, r r- intT1T2r5 r7 r: r; r< r=
__classcell__r& r& r$ r' r s
r c s~ e Zd ZdZdddefeeef eeee e ege
f dd fdd
Z e ee
ef d fd d
Zeejd d d
Z ZS )r a `Sintel `_ Dataset for optical flow.
The dataset is expected to have the following structure: ::
root
Sintel
testing
clean
scene_1
scene_2
...
final
scene_1
scene_2
...
training
clean
scene_1
scene_2
...
final
scene_1
scene_2
...
flow
scene_1
scene_2
...
Args:
root (str or ``pathlib.Path``): Root directory of the Sintel Dataset.
split (string, optional): The dataset split, either "train" (default) or "test"
pass_name (string, optional): The pass to use, either "clean" (default), "final", or "both". See link above for
details on the different passes.
transforms (callable, optional): A function/transform that takes in
``img1, img2, flow, valid_flow_mask`` and returns a transformed version.
``valid_flow_mask`` is expected for consistency with other datasets which
return a built-in valid mask, such as :class:`~torchvision.datasets.KittiFlow`.
loader (callable, optional): A function to load an image given its path.
By default, it uses PIL as its image loader, but users could also pass in
``torchvision.io.decode_image`` for decoding image data into tensors directly.
traincleanN)r split pass_namer r r c
s t j|||d t|ddd t|ddd |dkr>dd gn|g}t|d
}|d d }|D ]}|d
krpd n|}|| | } t| D ]|}
ttt| |
d } t t
| d D ]$} | j | | | | d gg7 _ q|d
kr| j ttt||
d 7 _ qq`d S )Nr r r rH rF testZ valid_valuesrI rG finalbothrP rG rO r Ztrainingr3 rF *.pngr *.flo)
r r r r oslistdirsortedr rA ranger6 r! r )
r# r rH rI r r passesZ flow_rootZ split_dirZ
image_rootZsceneZ
image_listir$ r& r' r s "zSintel.__init__r. c s t |S a Return example at given index.
Args:
index(int): The index of the example to retrieve
Returns:
tuple: A 3-tuple with ``(img1, img2, flow)``.
The flow is a numpy array of shape (2, H, W) and the images are PIL images.
``flow`` is None if ``split="test"``.
If a valid flow mask is generated within the ``transforms`` parameter,
a 4-tuple with ``(img1, img2, flow, valid_flow_mask)`` is returned.
r r5 r# r/ r$ r& r' r5 s
zSintel.__getitem__r( c C s t |S r* _read_flor+ r& r& r' r- s zSintel._read_flowr> r? r@ __doc__r r
rA r r r r r rB rC rD r5 npndarrayr- rE r& r& r$ r' r X s .
r c s e Zd ZdZdZddefeeef ee e
e
ege f dd fdd
Z e
eeef d fd d
Zeeejejf d d d
Z ZS )r a `KITTI `__ dataset for optical flow (2015).
The dataset is expected to have the following structure: ::
root
KittiFlow
testing
image_2
training
image_2
flow_occ
Args:
root (str or ``pathlib.Path``): Root directory of the KittiFlow Dataset.
split (string, optional): The dataset split, either "train" (default) or "test"
transforms (callable, optional): A function/transform that takes in
``img1, img2, flow, valid_flow_mask`` and returns a transformed version.
loader (callable, optional): A function to load an image given its path.
By default, it uses PIL as its image loader, but users could also pass in
``torchvision.io.decode_image`` for decoding image data into tensors directly.
TrF Nr rH r r r c s t j|||d t|ddd t|d |d }ttt|d d }ttt|d d }|rl|sttd
t||D ] \}}| j ||gg7 _ q~|d krttt|d d | _
d S )
NrJ rH rK rM r Zingimage_2z*_10.pngz*_11.pngzZCould not find the Kitti flow images. Please make sure the directory structure is correct.rF flow_occ) r r r r rU r rA FileNotFoundErrorzipr! r ) r# r rH r r images1images2r1 r2 r$ r& r' r s zKittiFlow.__init__r. c s t |S )a Return example at given index.
Args:
index(int): The index of the example to retrieve
Returns:
tuple: A 4-tuple with ``(img1, img2, flow, valid_flow_mask)``
where ``valid_flow_mask`` is a numpy boolean mask of shape (H, W)
indicating which flow values are valid. The flow is a numpy array of
shape (2, H, W) and the images are PIL images. ``flow`` and ``valid_flow_mask`` are None if
``split="test"``.
rZ r[ r$ r& r' r5 s
zKittiFlow.__getitem__r( c C s t |S r* )_read_16bits_png_with_flow_and_valid_maskr+ r& r& r' r- s zKittiFlow._read_flow)r> r? r@ r_ r0 r r
rA r r r r r rB rC rD r5 tupler` ra r- rE r& r& r$ r' r s
r c sh e Zd ZdZd
eeef eee dd fdd
Z e
ee e f d fdd Z
eejd
d d Z ZS )r a `FlyingChairs `_ Dataset for optical flow.
You will also need to download the FlyingChairs_train_val.txt file from the dataset page.
The dataset is expected to have the following structure: ::
root
FlyingChairs
data
00001_flow.flo
00001_img1.ppm
00001_img2.ppm
...
FlyingChairs_train_val.txt
Args:
root (str or ``pathlib.Path``): Root directory of the FlyingChairs Dataset.
split (string, optional): The dataset split, either "train" (default) or "val"
transforms (callable, optional): A function/transform that takes in
``img1, img2, flow, valid_flow_mask`` and returns a transformed version.
``valid_flow_mask`` is expected for consistency with other datasets which
return a built-in valid mask, such as :class:`~torchvision.datasets.KittiFlow`.
rF N)r rH r r c
s t j||d t|ddd t|d }ttt|d d }ttt|d d }d }tj || svt
d
t j t|| t j
d }tt|D ]h}|| } |d kr| d
ks|dkr| dkr| j|| g7 _| j|d| |d| d
gg7 _qd S )N)r r rH )rF valrM r r; z*.ppmrR zFlyingChairs_train_val.txtzmThe FlyingChairs_train_val.txt file was not found - please download it from the dataset page (see docstring).)ZdtyperF r rl r )r r r r rU r rA rS pathexistsre r` Zloadtxtint32rV r6 r r! )
r# r rH r imagesflowsZsplit_file_nameZ
split_listrX Zsplit_idr$ r& r' r s zFlyingChairs.__init__r. c s t |S )a Return example at given index.
Args:
index(int): The index of the example to retrieve
Returns:
tuple: A 3-tuple with ``(img1, img2, flow)``.
The flow is a numpy array of shape (2, H, W) and the images are PIL images.
``flow`` is None if ``split="val"``.
If a valid flow mask is generated within the ``transforms`` parameter,
a 4-tuple with ``(img1, img2, flow, valid_flow_mask)`` is returned.
rZ r[ r$ r& r' r5 * s
zFlyingChairs.__getitem__r( c C s t |S r* r\ r+ r& r& r' r- 9 s zFlyingChairs._read_flow)rF N)r> r? r@ r_ r
rA r r r r rB rC rD r5 r` ra r- rE r& r& r$ r' r s & r c s e Zd ZdZddddefeeef eeeee e ege
f dd fdd
Z e ee
ef d fd
d Zeejd d
dZ ZS )r a `FlyingThings3D `_ dataset for optical flow.
The dataset is expected to have the following structure: ::
root
FlyingThings3D
frames_cleanpass
TEST
TRAIN
frames_finalpass
TEST
TRAIN
optical_flow
TEST
TRAIN
Args:
root (str or ``pathlib.Path``): Root directory of the intel FlyingThings3D Dataset.
split (string, optional): The dataset split, either "train" (default) or "test"
pass_name (string, optional): The pass to use, either "clean" (default) or "final" or "both". See link above for
details on the different passes.
camera (string, optional): Which camera to return images from. Can be either "left" (default) or "right" or "both".
transforms (callable, optional): A function/transform that takes in
``img1, img2, flow, valid_flow_mask`` and returns a transformed version.
``valid_flow_mask`` is expected for consistency with other datasets which
return a built-in valid mask, such as :class:`~torchvision.datasets.KittiFlow`.
loader (callable, optional): A function to load an image given its path.
By default, it uses PIL as its image loader, but users could also pass in
``torchvision.io.decode_image`` for decoding image data into tensors directly.
rF rG leftN)r rH rI camerar r r c s t j|||d t|ddd | }t|ddd dgdgddgd| }t d d
d d krld d
gn g}t|d }d} t||| D ]\\} ttt || | d }
t fdd|
D }
ttt |d | d } t fdd| D } |
r| st
dt |
| D ]\} }
ttt | d }ttt |
d }t t
|d D ]}dkr| j|| ||d gg7 _| j|| g7 _nBdkr`| j||d || gg7 _| j||d g7 _q`q qd S )NrJ rH rK rM rI rN Zframes_cleanpassZframes_finalpassrs )rr rightrP rP rr rt r ) into_future into_pastz*/*c 3 s | ]}t | V qd S r* r ).0 image_dir)rs r& r' z z*FlyingThings3D.__init__..Z optical_flowc 3 s | ]}t | V qd S r* r )rw flow_dirrs directionr& r' ry } rz zcCould not find the FlyingThings3D flow images. Please make sure the directory structure is correct.rQ z*.pfmr ru rv )r r r upperr itertoolsproductrU r rA re rf rV r6 r! r )r# r rH rI rs r r rW ZcamerasZ
directionsZ
image_dirsZ flow_dirsrx r{ rp rq rX r$ r| r' r ] sB
zFlyingThings3D.__init__r. c s t |S rY rZ r[ r$ r& r' r5 s
zFlyingThings3D.__getitem__r( c C s t |S r* )r r+ r& r& r' r- s zFlyingThings3D._read_flowr^ r& r& r$ r' r = s" "
3 r c s e Zd ZdZdZddefeeef ee e
e
ege f dd fdd
Z ee
ejejf dd d
Zeeeef d fd d
Z ZS )r a `HD1K `__ dataset for optical flow.
The dataset is expected to have the following structure: ::
root
hd1k
hd1k_challenge
image_2
hd1k_flow_gt
flow_occ
hd1k_input
image_2
Args:
root (str or ``pathlib.Path``): Root directory of the HD1K Dataset.
split (string, optional): The dataset split, either "train" (default) or "test"
transforms (callable, optional): A function/transform that takes in
``img1, img2, flow, valid_flow_mask`` and returns a transformed version.
loader (callable, optional): A function to load an image given its path.
By default, it uses PIL as its image loader, but users could also pass in
``torchvision.io.decode_image`` for decoding image data into tensors directly.
TrF Nrb c
sL t j|||d t|ddd t|d }|dkrtdD ]}ttt|d d |d
d }ttt|d d
|d
d }tt|d D ]8}| j || g7 _ | j
|| ||d gg7 _
qq r? r@ r_ r0 r r
rA r r r r r rk r` ra r- rB rC rD r5 rE r& r& r$ r' r s
r r( c C s t | d}tj|ddd }|dkr0tdttj|ddd}ttj|ddd}tj|d d
| | d}|||d
d
d dW d S 1 s0 Y d S )
z#Read .flo file in Middlebury formatrbc )counts PIEHz)Magic number incorrect. Invalid .flo filez s. :[FDfL