What it does

This is a master’s thesis project for video object tracking in TensorFlow, built to compete in ImageNet’s VID challenge. It chains together existing open-source implementations: YOLO or TensorBox for detection, hand-rolled post-processing for temporal smoothing, and Inception for classification. You feed it a video file; it spits out an annotated MP4.

The interesting bit

The architecture explicitly copies the T-CNN paper’s cascade strategy—detection first, then temporal information, then context—but replaces the trainable temporal components with non-trainable post-processing algorithms in Utils_Video.py. It’s a frankenstein that admits it’s a frankenstein.

Key highlights

• Supports YOLO (single-class detection) and TensorBox (multi-class) pipelines
• Includes dataset preprocessing scripts with brute-force and lightweight modes
• Provides trained weights for both Inception and TensorBox via Mega.nz links
• Hardcoded to 640×480 PNG for TensorBox; you’ll need to hack the resize scripts for other dimensions
• Requires OpenCV, TensorFlow, and Python; installation guide included

Caveats

• Temporal information is “retrieved through some post processing algorithm… NOT TRAINABLE” — the README’s own caps
• Weight download links are from 2017 (Mega.nz); longevity unclear
• Early output had frame-ordering bugs causing flicker; author notes this was “then solved” but the fix isn’t detailed
• “I will soon put a weight file to download” — still pending as of last update 10-03-2017

Verdict

Worth a look if you’re studying how to bolt together 2016-era detection models for video, or if you need a concrete (if dated) reference for the ImageNet VID pipeline. Skip it if you want a maintained, trainable end-to-end tracker—this is explicitly thesis-grade glue code with the scaffolding still showing.