Reid result is empty

[chenjx1005]

Thank you for your sharing. When I try to reproduce your result, I find two problems .

1 .the reid_dict in https://github.com/KevinQian97/ELECTRICITY-MTMC/blob/master/identifier/test.py#L66 is a empty dict. That is the final result is just the single camera tracking result without vehicle_id in rm_dict.

$python ./identifier/test.py --test-batch-size 256 --test_set aic_test --use-avai-gpus --load-weights ./models/resnet101-Aic/model.pth.tar-9
Currently using GPU 0
Initializing image data manager
=> Initializing TEST datasets
Initializing model: resnet101
Model size: 42.500 M
Successfully loaded pretrained weights from "./models/resnet101-Aic/model.pth.tar-9"
** The following layers are discarded due to unmatched keys or layer size: ['classifier.weight', 'classifier.bias']
Matching aic_test ...
Extracted features for query set, obtained 963-by-2048 matrix
Extracted features for gallery set, obtained 123698-by-2048 matrix
=> BatchTime(s)/BatchSize(img): 0.399/256
/opt/conda/conda-bld/pytorch_1587428266983/work/torch/csrc/utils/python_arg_parser.cpp:756: UserWarning: This overload of addmm_ is deprecated:
        addmm_(Number beta, Number alpha, Tensor mat1, Tensor mat2)
Consider using one of the following signatures instead:
        addmm_(Tensor mat1, Tensor mat2, *, Number beta, Number alpha)
{45: {145: True, 162: True, 154: True, 150: True, 230: True, 350: True, 10: True, 64: True, 237: True, 414: True, 81: True, 33: True, 97: True, 23: True, 9: True, 309: True, 259: True}, 46: {42: True, 133: True, 468: True, 68: True, 66: True, 106: True, 9: True, 54: True, 141: True}, 42: {34: True, 646: True, 144: True, 1247: True, 550: True, 1033: True, 1291: True, 286: True, 322: True, 538: True, 1294: True, 541: True, 537: True, 2: True, 589: True, 587: True, 508: True, 542: True, 395: True, 1227: True}, 44: {20: True, 353: True, 14: True, 39: True, 499: True, 232: True, 123: True, 489: True}, 41: {155: True, 1240: True, 949: True, 278: True, 11: True, 1021: True, 1054: True, 372: True, 271: True, 921: True, 1183: True, 257: True}, 43: {494: True, 108: True, 53: True, 464: True, 129: True}} {}

2. as the issue https://github.com/KevinQian97/ELECTRICITY-MTMC/issues/5 posted, is such many duplicated vehicles in a same frame of a same video a normal case? tracklets.txt has 424146 lines. But after execute python ./identifier/preprocess/extract_img.py there are only 124661 images left in exp/imgs/aic_test. Other vehicles images are not considered at all because of overwritting .

[KevinQian97]

Hi chenjx,

Thanks for your interest! For question one, according to our implementation and experiments of others, the Reid result is not empty. Would u mind providing more information about your results? i.e. you directly run the test bash with our provided model or use the model trained by yourself. Also, would u mind providing the track3.txt generated here? For question two, there are some redundant vehicle ids, but I didn't remember that number is so large. Anyway, for the official metric, redundant id won't affect the final performance.

[KevinQian97]

Sry, just notice your printed results. The first dict of the output is the re-id dict and the second one is the filtering dict.

[dengxyking]

---

reid_dict,rm_dict = calc_reid(result) print(rm_dict,reid_dict)

---

rm_dict is the output is the re-id dict and reid_dict is the filtering dict?

[chenjx1005]

Sry, just notice your printed results. The first dict of the output is the re-id dict and the second one is the filtering dict.

No. The first dict of the output is the remove dict and the second one is the re-id dict.
the reid dict should look like 770: {'dis': 11.902046, 'id': 2638}, 365: {'dis': 11.598312, 'id': 2639}, 437: {'dis': 10.53886, 'id': 2640}, 164: {'dis': 9.796799, 'id': 2642}

[Hlings]

@chenjx1005 I meet the same problem. Have you solved this? Thanks.

[chenjx1005]

@Hlings No. I give up using this project. The result of this solution has many repeated lines, which causes the evaluation code gives a much higher score than it should have. This is a bug of the AIC20 and has been fixed in AIC21.