Replaced every instance of np.float with built-in float #399
Conversation
…Tracker update method
yolox/tracker/byte_tracker.py
Outdated
| @@ -143,20 +150,29 @@ def __repr__(self): | |||
|
|
|||
|
|
|||
| class BYTETracker(object): | |||
| def __init__(self, args, frame_rate=30): | |||
| def __init__(self, args, feature_extractor, global_stracks:List[STrack]=[], frame_rate=30):###<----- added feature_extractor & global_stracks attribute to constructor | |||
There was a problem hiding this comment.
if I'm not wrong, feature_extractor should be feature extraction model right?
There was a problem hiding this comment.
Feature Extractor in BYTETracker
Yes, the feature_extractor is passed as an argument when initializing BYTETracker. This extractor is responsible for generating features from detected objects, which are then used for tracking purposes.
The specific FeatureExtractor class I use comes from the deep-person-reid repository:
https://github.com/KaiyangZhou/deep-person-reid/blob/master/torchreid/utils/feature_extractor.py
There was a problem hiding this comment.
Feature Extraction in ByteTrack's Update Method
The update method in the original ByteTrack repository (https://github.com/ifzhang/ByteTrack.git) does not employ feature extraction for tracking. Instead, it relies solely on a Kalman filter to predict the position of bounding boxes in the current frame based on the position and velocity data associated with objects in the previous frame.
However, this approach has a drawback: the Kalman filter cannot predict the box for a person who has been lost for more than a certain number of frames (30 by default). Consequently, if a person disappears for more than 30 frames, the tracker simply ignores them, and if they reappear, they are assigned a new ID. This is problematic as it fails to recognize the reappearance of the same individual.
To address this issue, I propose storing the features of every unique person detected so far. This way, if a person disappears for more than the buffer size (30 frames by default) and then reappears, we can compare their features with the previously stored ones to determine whether it's the same person or not. This approach allows for more accurate tracking and avoids assigning new IDs to individuals who have simply been out of sight for a short period.
yolox/tracker/byte_tracker.py
Outdated
|
|
||
| def update(self, output_results, img_info, img_size): | ||
| def update(self, frame, output_results, img_info, img_size): ###<----- added frame attribute to update method |
There was a problem hiding this comment.
it would be fine to me, but maybe need documented a bit? might be confused to others?
There was a problem hiding this comment.
yeah, you are right. I have used the README.md file from the original ByteTracker repo as is, and that could be confusing. I will modify it to specify what changes I have made and how it is different from the original repository.
There was a problem hiding this comment.
This code snippet utilizes a modified update() method that takes a frame as an argument to extract features for each individual detected in the image. Here's a breakdown of the process:
Step 1: Person Detection
The code uses YOLOv8 (or any other object detector) to detect individuals in the frame.
YOLOv8 returns a result containing bounding boxes, class IDs, confidence scores, and other information for each detected object.
Step 2: Tracking with BYTETracker
The detected bounding boxes are passed to the BYTETracker class along with the frame.
BYTETracker is responsible for tracking individual objects across multiple frames.
Step 3: Region Cropping
The code crops the regions specified by the bounding boxes from the frame.
These cropped regions represent the individual persons detected in the image.
Step 4: Feature Extraction
The cropped regions are passed as a list to a feature_extractor function.
This function extracts features from each individual region, such as facial features or body posture.
Reason for Taking frame as Input:
The modified update() method takes the frame as input to enable feature extraction for each individual detected in the image. This allows for more accurate tracking and identification of individuals across multiple frames.
There was a problem hiding this comment.
Love the changes! I was working on ByteTrack with DETR and Yolov8 before, even with ReID model, the results seem still bad as I want to get more accurate results for unique counts. Hope this improve the results!
… removed the elements from u_detection for which there was a matching track in removed_stracks
Hey there! I've just replaced every instance of
np.floatwith the built-infloat. Thenp.floathas been deprecated.