search

chengche6230 / ReST

[ICCV 2023] ReST: A Reconfigurable Spatial-Temporal Graph Model for Multi-Camera Multi-Object Tracking
MIT License
130 stars 15 forks source link

Non-replicable campus inference #14

Open iflab-felipe opened 2 months ago

iflab-felipe commented 2 months ago

Hi. I trained my TG and used your pretrained SG (for CAMPUS Garden 1) as another issue suggested (#6) and I cannot replicate your results. I am using osnet_x1_0 as my ReID model. IDs are changing constantly for every frame and person, even though the spatial graph seems ok, as if the SG was working but not the TG. In this sense I also have to say that the results when using both of your pretrained weights is worse than this mixed approach somehow. Results obtained along with an iteration example and the SG:

2024-04-12 13:02:16.769 | INFO | src.datasets.dataset:getInference:312 - Loaded graph with 39 nodes and 1108 edges. 2024-04-12 13:02:16.802 | INFO | src.utils.tracklet:postprocessing_sg:133 - Post-processing done, removed 5 edges. Graph: 39 nodes, 30 edges. 2024-04-12 13:02:17.011 | INFO | src.tracker:_test_one_epoch:233 - Iteration 565: Spatial Graph done.SG: 16 nodes and 0 edges. 2024-04-12 13:02:17.036 | INFO | src.utils.tracklet:postprocessing_tg:324 - Post-processing done, removed 1 edges. Graph: 35 nodes, 13 edges. 2024-04-12 13:02:17.324 | INFO | src.tracker:_test_one_epoch:261 - Iteration 565: Temporal Graph done.TG: 22 nodes and 0 edges. 2024-04-12 13:02:17.324 | INFO | src.tracker:_test_one_epoch:264 - 【Finished inference iteration 565/568】

IDF1 IDP IDR Rcll Prcn GT MT PT ML FP FN IDs FM MOTA MOTP IDt IDa IDm 0 25.8% 31.5% 25.8% 99.9% 99.9% 15 15 0 0 4 4 1436 4 70.6% 0.012 584 596 1 1 30.4% 38.0% 30.4% 100.0% 100.0% 12 12 0 0 0 0 1051 0 77.8% 0.000 355 489 2 2 21.4% 27.6% 21.4% 100.0% 100.0% 15 15 0 0 0 0 2145 0 73.1% 0.003 1012 836 3 3 22.5% 28.5% 22.5% 100.0% 100.0% 13 13 0 0 0 0 1216 0 69.4% 0.016 615 550 5 OVERALL 24.6% 31.0% 24.6% 100.0% 100.0% 55 55 0 0 4 4 5848 4 72.9% 0.007 2566 2471 11

cluster_12

I would kindly ask for some advice to find what is going wrong here. Thank you so much.

chengche6230 commented 2 months ago

Hi,

Can you provide me more details, e.g. training loss, iteration, etc., about your training process? I didn't encounter such issue if the graph is well-trained.