This repository contains the description of my Research & Development project.
The target of this project was to categorize the contemporary methods developed for object detection in videos and conduct a comparison and analyse whether the inclusion of temporal information has any benefits to the detection process or not. The comparison was conducted by implementing two methods namely: Framewise approach (object detection was conducted on every single frame of the input video) and Coupled approach (object detection was conducted on sparse frames and Open Loop Kalman Filter was applied on the remaining frames of the input video).
The repository has the following structure:
(i) contemporary research papers
- per frame methodologies
- temporal information methodologies
(ii) custom data
- decomposed_input_video: contains the frames comprising the input videos
- ground_truth_annotation: manually annotated frames obtained using annotation.py
- input_video: contains the test input video
(iii) output
- decomposed_output_video:
- coupled: contains frames decomposed from the output video of coupled approach
- framewise: contains frames decomposed from the output video of framewise approach
- coupled: contains frames decomposed from the output video of coupled approach
- detected_coordinates.txt: contains the bounding boxes coordinates of detected approach object for both approaches
- fps_count: frame per second count readings for both methods
- inference_times: contains the timings required for both approaches to process the video
- output_video: contains the video output of both approaches
- iou_scissor_om.png: comparison of IoU for both approaches for a given input
- it_scissor_om.png: comparison of inference times for both approaches for a given input
(iv) src
- annotation.py: script to annotate the video inputs and generate annotated frames
- coupled_dt_mobile.ipynb: pipeline implementing Coupled approach
- Fragmenting_videos_to_frames.py: decomposing video inputs to constituent frames
- framewise_mobile.ipynb: pipeline implementing Framewise approach
- inference_time_comparison.py: script to compare the inference times for both approach and generate a plot
- iou_calculator.py: script to compare the IoU for both approaches with ground truth and generate a plot
- object_detection_tutorial_video_webcam.py: test webcam working with tensorflow api