A more realtime adaptation of Deep SORT.
Adapted from the official repo of Simple Online and Realtime Tracking with a Deep Association Metric (Deep SORT)
See their paper for more technical information.
requirements.txt
gives the default packages required (it installs torch/torchvision to use the default mobilenet embedder), modify accordingly.
Main dependencies are:
- Python3
- NumPy,
pip install numpy
- SciPy,
pip install scipy
- cv2,
pip install opencv-python
- (optional) Embedder requires Pytorch & Torchvision (for the default MobiletnetV2 embedder) or Tensorflow 2+
pip install torch torchvision
pip install tensorflow
- (optional) Additionally, to use Torchreid embedder,
torchreid
Python package needs to be installed. You can follow installation guide onTorchreid
's page. Without using conda, you can simply clone that repository and do apython3 -m pip install .
from inside the repo. - (optional) To use CLIP embedder,
pip install git+https://github.com/openai/CLIP.git
- from PyPI via
pip3 install deep-sort-realtime
orpython3 -m pip install deep-sort-realtime
- or, clone this repo & install deep-sort-realtime as a python package using
pip
or as an editable package if you like (-e
flag)
cd deep_sort_realtime && pip3 install .
- or, download
.whl
file in this repo's releases
Example usage:
from deep_sort_realtime.deepsort_tracker import DeepSort
tracker = DeepSort(max_age=5)
bbs = object_detector.detect(frame)
tracks = tracker.update_tracks(bbs, frame=frame) # bbs expected to be a list of detections, each in tuples of ( [left,top,w,h], confidence, detection_class )
for track in tracks:
if not track.is_confirmed():
continue
track_id = track.track_id
ltrb = track.to_ltrb()
-
To add project-specific logic into the
Track
class, you can make a subclass (ofTrack
) and pass it in (override_track_class
argument) when instantiatingDeepSort
. -
Example with your own embedder/ReID model:
from deep_sort_realtime.deepsort_tracker import DeepSort
tracker = DeepSort(max_age=5)
bbs = object_detector.detect(frame) # your own object detection
object_chips = chipper(frame, bbs) # your own logic to crop frame based on bbox values
embeds = embedder(object_chips) # your own embedder to take in the cropped object chips, and output feature vectors
tracks = tracker.update_tracks(bbs, embeds=embeds) # bbs expected to be a list of detections, each in tuples of ( [left,top,w,h], confidence, detection_class ), also, no need to give frame as your chips has already been embedded
for track in tracks:
if not track.is_confirmed():
continue
track_id = track.track_id
ltrb = track.to_ltrb()
- Look into
deep_sort_realtime/deepsort_tracker.py
for more hyperparameters which you can tune to suit your use-case.
The original Track.to_*
methods for retrieving bounding box values returns only the Kalman predicted values. However, in some applications, it is better to return the bb values of the original detections the track was associated to at the current round.
Here we added an orig
argument to all the Track.to_*
methods. If orig
is flagged as True
and this track is associated to a detection this update round, then the bounding box values returned by the method will be that associated to the original detection. Otherwise, it will still return the Kalman predicted values.
orig_strict
argument in all the Track.to_*
methods is only active when orig
is True
. Flagging orig_strict=True
will mean it will output None
when there's no original detection associated to this track at current frame, otherwise normally it will return Kalman predicted values.
Supplementary info can be pass into the track from the detection. Detection
class now has an others
argument to store this and pass it to the associate track during update. Can be retrieved through Track.get_det_supplementary
method. Can be passed in through others
argument of DeepSort.update_tracks
, expects to be a list with same length as raw_detections
. Examples of when you will this includes passing in corresponding instance segmentation masks, to be consumed when iterating through the tracks output.
Other than horizontal bounding boxes, detections can now be given as polygons. We do not track polygon points per se, but merely convert the polygon to its bounding rectangle for tracking. That said, if embedding is enabled, the embedder works on the crop around the bounding rectangle, with area not covered by the polygon masked away.
When instantiating a DeepSort
object (as in deepsort_tracker.py
), polygon
argument should be flagged to True
. See DeepSort.update_tracks
docstring for details on the polygon format. In polygon mode, the original polygon coordinates are passed to the associated track through the supplementary info.
-
Remove "academic style" offline processing style and implemented it to take in real-time detections and output accordingly.
-
Provides both options of using an in-built appearance feature embedder or to provide embeddings during update
-
Added pytorch mobilenetv2 as appearance embedder (tensorflow embedder is also available now too).
-
Added CLIP network from OpenAI as embedder (pytorch).
-
Skip nms completely in preprocessing detections if
nms_max_overlap == 1.0
(which is the default), in the original repo, nms will still be done even if threshold is set to 1.0 (probably because it was not optimised for speed). -
Now able to override the
Track
class with a custom Track class (that inherits fromTrack
class) for custom track logic -
Takes in today's date now, which provides date for track naming and facilities track id reset every day, preventing overflow and overly large track ids when system runs for a long time.
from datetime import datetime today = datetime.now().date()
-
Now supports polygon detections. We do not track polygon points per se, but merely convert the polygon to its bounding rectangle for tracking. That said, if embedding is enabled, the embedder works on the crop around the bounding rectangle, with area not covered by the polygon masked away. Read more here.
-
The original
Track.to_*
methods for retrieving bounding box values returns only the Kalman predicted values. In some applications, it is better to return the bb values of the original detections the track was associated to at the current round. Added aorig
argument which can be flaggedTrue
to get that. Read more here. -
Added
get_det_supplementary
method toTrack
class, in order to pass detection related info through the track. Read more here. -
[As of 2fad967] Supports background masking by giving instance mask to
DeepSort.update_tracks
. Read more here. -
Other minor adjustments/optimisation of code.
In package deep_sort
is the main tracking code:
detection.py
: Detection base class.kalman_filter.py
: A Kalman filter implementation and concrete parametrization for image space filtering.linear_assignment.py
: This module contains code for min cost matching and the matching cascade.iou_matching.py
: This module contains the IOU matching metric.nn_matching.py
: A module for a nearest neighbor matching metric.track.py
: The track class contains single-target track data such as Kalman state, number of hits, misses, hit streak, associated feature vectors, etc.tracker.py
: This is the multi-target tracker class.
python3 -m unittest
Default embedder is a pytorch MobilenetV2 (trained on Imagenet).
For convenience (I know it's not exactly best practice) & since the weights file is quite small, it is pushed in this github repo and will be installed to your Python environment when you install deep_sort_realtime.
Torchreid is a person re-identification library, and is supported here especially useful for extracting features of humans. Torchreid
will need to be installed (see dependencies section above) It provides a zoo of models. Select model type to use, note the model name and provide as arguments. Download the corresponding model weights file on the model zoo site and point to the downloaded file. Model 'osnet_ain_x1_0' with domain generalized training on (MS+D+C) is provide by default, together with the corresponding weights. If embedder='torchreid'
when initalizing DeepSort
object without specifying embedder_model_name
or embedder_wts
, it will default to that.
from deep_sort_realtime.deepsort_tracker import DeepSort
tracker = DeepSort(max_age=5, embedder='torchreid')
bbs = object_detector.detect(frame)
tracks = tracker.update_tracks(bbs, frame=frame) # bbs expected to be a list of detections, each in tuples of ( [left,top,w,h], confidence, detection_class )
for track in tracks:
if not track.is_confirmed():
continue
track_id = track.track_id
ltrb = track.to_ltrb()
CLIP is added as another option of embedder due to its proven flexibility and generalisability. Download the CLIP model weights you want at deep_sort_realtime/embedder/weights/download_clip_wts.sh and store the weights at that directory as well, or you can provide your own CLIP weights through embedder_wts
argument of the DeepSort
object.
Available now at deep_sort_realtime/embedder/embedder_tf.py
, as alternative to (the default) pytorch embedder. Tested on Tensorflow 2.3.1. You need to make your own code change to use it.
The tf MobilenetV2 weights (pretrained on imagenet) are not available in this github repo (unlike the torch one). Download mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_224.h5
from https://github.com/JonathanCMitchell/mobilenet_v2_keras/releases/tag/v1.1. You may drop it into deep_sort_realtime/embedder/weights/
before pip installing.
If instance mask is given during DeepSort.update_tracks
with no external appearance embeddings given, the mask will be used to mask out the background of the corresponding detection crop so that only foreground information goes into the embedder. This reduces background bias.
Example cosine distances between images in ./test/
("diff": rock vs smallapple, "close": smallapple vs smallapple slightly augmented)
.Testing pytorch embedder
close: 0.012196660041809082 vs diff: 0.4409685730934143
.Testing Torchreid embedder
Model: osnet_ain_x1_0
- params: 2,193,616
- flops: 978,878,352
Successfully loaded pretrained weights from "/Users/levan/Workspace/deep_sort_realtime/deep_sort_realtime/embedder/weights/osnet_ain_ms_d_c_wtsonly.pth"
close: 0.012312591075897217 vs diff: 0.4590487480163574