tmoopenn / seq-nms

Implementation of the seq-nms post-processing algorithm for video object detection

Seq-NMS for Video Object Detection

This project provides a python implementation of [Seq-NMS] (https://arxiv.org/abs/1602.08465) post-processing algorithm for video object detection.

Setup Details

This repository is developed and tested with python 3.7.3. Setup requires and is tested with cython (0.29.20) and numpy (1.18.5)

Run the following command to build and compile the cython file compute_overlap.pyx:

python setup.py build_ext --inplace

seq-nms's Issues

Integrating seq-nms with yolov5s model

Hi @tmoopenn , can you guide how and where to use this seq-nms repo on yolov5. I tried using the seq_nms function on ultralytics/yolov5 repo but function parameters don't correspond well.

Output Interpretation

Hi,

I can't understand the output of this script. So I have a sequence of 22 frames and 4 bboxes for each frame with different scores.
and here's the output of the script:

BOX GRAPH SHAPE (21, 4)
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] 18.8158267611
[1] 0.996027112

what does this mean?

box_graph shape

box_graph is list of shape (num_frames - 1, num_boxes, k). What is the meaning of k? Could you provide a example boxes and scores for boxes = np.load('/path/to/boxes') and cores = np.load('/path/to/scores')#(num_frames, num_boxes). Thank you.

Could you please clarify the license of this repository?

Hi,

I was wondering if you could clarify the license of this repository?
I am currently working on a task for video image recognition and would like to use your seq-nms implementation.
Thank you very much for your time.

Questions on the result

Thank you for excellent codes, but I want to ask you something about the outputs. I have the result:
BOX GRAPH SHAPE (463, 300)
[240, 248, 246, ...] 65.35569
[204, 202, ...] 4.000431
[249] 0.19185601
Of course the length of the first list is 464... My question is, isn't the first list enough to be an output? I can draw bboxes on every frames with it, no need to suppress the best_sequence to length 1. Or if it is not, how can I make use of the last list [249] to calculate mAP of this video clip?

How to use this code on the top of yolo prediction?

Hi @tmoopenn I want to use this algorithm to reinforce my yolo prediction. Can you give me the example of boxes and scores file how it looks like that you load for the testing. I will create these files in your format