Matchnet-For-VOT
https://github.com/hanxf/matchnet adapted for VOT dataset
Files have been edited to run without .sh files. This is for the purpose of testing only. Python files of:
1. generate_patch_dp.py: run this to generate yosemite dataset
- to reduce time to experiment, run for only one dataset.
- Procedure to run for other datasets can be found here: https://github.com/hanxf/matchnet
- Trained models can be downloaded from here: https://github.com/hanxf/matchnet/tree/master/models
2. evaluate_matchnet.py: run this to train and test for yosemite only.
- the model works well for grayscale data
- key take aways would be: Matching is invariant to illuminations (not scale)
3. eval_matchnetVOT2016.py: run this to compare current VOT 2016 dataset current frame with the previous frame.
- This works fine too. Data templates are choses randomly.
- The above point adds to the error in detection; in case templates are not choses around the grouth truth object points.
- would work better if RGB - all 3 color channels were considered. Imples network has to be modified.
4. eval_VOTFirst Frame.py: run this to compare current VOT 2016 dataset current frame with the first frame.
- question to be answered: how does the matchnet compare to object variations?
- problem: matchnet is not trained for objects, but for tetural data.
- solution: train on object data, then look further into how network adapts to object variations.
results: Top Match and Matches above 0.95 score (brighter red implies higher score)
Purpose
Testing Matchnet (courtesy - Xufeng Han)