How can I make filtered_detections?

[songheony]

I'd like to test this tracker on my custom dataset. How can I make filtered_detections?

[derkbreeze]

You can perform tracktor's tracking on raw sequences, and use their output unlabeled detections as filtered detections

[songheony]

As a result of using the tracking result of the tracktor (https://github.com/phil-bergmann/tracking_wo_bnw) as a detection according to your answer, I've got a much better performance than what is written in your paper. Is this just a bug?

MOT16-10 74.6% 91.7% 62.9% 65.7% 95.9% 54 21 25 8 345 4220 25 85 62.7% 0.177 16 15 10 MOT16-05 69.1% 90.4% 56.0% 58.2% 94.0% 125 29 64 32 254 2848 22 58 54.2% 0.171 23 11 14 MOT16-09 77.4% 94.3% 65.6% 67.6% 97.2% 25 12 12 1 103 1701 10 35 65.5% 0.139 6 3 2 MOT16-11 69.0% 83.4% 58.8% 66.1% 93.8% 69 20 32 17 402 3108 15 28 61.6% 0.135 7 12 4 MOT16-02 54.6% 95.1% 38.3% 39.7% 98.6% 54 7 26 21 101 10747 8 71 39.1% 0.138 9 1 4 MOT16-04 75.2% 94.0% 62.6% 65.7% 98.5% 83 31 29 23 462 16316 10 54 64.7% 0.134 9 5 6 MOT16-13 66.0% 91.2% 51.7% 55.0% 96.9% 107 34 42 31 201 5158 18 48 53.0% 0.156 22 7 13 OVERALL 70.5% 92.3% 57.0% 60.1% 97.3% 517 154 230 133 1868 44098 108 379 58.3% 0.144 92 54 53

[yinjunbo]

@songheony, my apologies for the late reply.
Here is the code for generating our filtered detections (https://github.com/otakudj/ADSS/blob/master/code/metric_v2.m#L91-L152). This just provides an alternative in comparison with simple NMS or confidence-based filtering methods. Correct initialization for new targets is of crucial importance in SOT-based paradigm, and our method can benefit a lot from better detections. As suggest by @derkbreeze, you can just treat other MOT tracker's results as detections to initialize our SOT-based tracker. Hence your results may be reasonable based on the Tracktor's powerful results.