Non-replicable Evaluation on PETS09-S2L1

[mkg1204]

ReST is really a nice work, and I'm very intersting about it. But when I trying to evaluate it on the PETS09-S2L1 dataset using the provided checkpoints in this repository, I got the following results, which is different from the results reported in your paper. The camera number in the metainfo.json is four, so I want to know which four cameras has been used in this experiments?

         IDF1   IDP   IDR   Rcll   Prcn GT MT PT ML FP FN IDs  FM  MOTA  MOTP IDt IDa IDm 
0       67.2% 67.9% 67.2% 100.0% 100.0%  7  7  0  0  0  0  44   0 88.8% 0.005   6  26   0 
1       35.4% 36.4% 35.4% 100.0% 100.0%  7  7  0  0  0  0  76   0 76.6% 0.000   9  50   0 
2       28.0% 39.2% 28.0% 100.0% 100.0%  7  7  0  0  0  0 171   0 63.8% 0.004  90  70   0 
3       32.6% 42.5% 32.6% 100.0% 100.0%  6  6  0  0  0  0 113   0 62.8% 0.000  49  48   1 
OVERALL 40.8% 47.8% 40.8% 100.0% 100.0% 27 27  0  0  0  0 404   0 73.0% 0.002 154 194   1 

Besides, as shown in the metainfo.json, 90% frames are used to train the models on CAMPUS and PETS09-S2L1 datasets, which is different with other works like LMGP. Why those two datasets are splited in this way?

Thanks for your nice work!

[chengche6230]

Hi,

We used the first four cameras in our experiment and followed the same setting in Wildtrack for consistency. Feel free to run different experiment settings.

Best

[jingliang0910]

hi，how can i download CAMPUS and PETS09-S2L1 datasets？I find their website can not be opened.