mirrors
/
deep-person-reid
mirror of https://github.com/KaiyangZhou/deep-person-reid.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
deep-person-reid/docs/MODEL_ZOO.md

11 KiB
Raw Blame History

Model Zoo

  • Results are presented in the format of <Rank-1 (mAP)>.
  • When computing model size and FLOPs, only layers that are used at test time are considered (see torchreid.utils.compute_model_complexity).
  • Asterisk (*) means the model is trained from scratch.
  • combineall=True means all images in the dataset are used for model training.
  • Why not use heavy data augmentation like random erasing for model training? It's because heavy data augmentation might harm the cross-dataset generalization performance (see this paper).

ImageNet pretrained models

Model Download
shufflenet model
mobilenetv2_x1_0 model
mobilenetv2_x1_4 model
mlfn model
osnet_x1_0 model
osnet_x0_75 model
osnet_x0_5 model
osnet_x0_25 model
osnet_ibn_x1_0 model
osnet_ain_x1_0 model
osnet_ain_x0_75 model
osnet_ain_x0_5 model
osnet_ain_x0_25 model

Same-domain ReID

Model # Param (10^6) GFLOPs Loss Input Transforms Distance market1501 dukemtmcreid msmt17
resnet50 23.5 2.7 softmax (256, 128) random_flip, random_crop euclidean 87.9 (70.4) 78.3 (58.9) 63.2 (33.9)
resnet50_fc512 24.6 4.1 softmax (256, 128) random_flip, random_crop euclidean 90.8 (75.3) 81.0 (64.0) 69.6 (38.4)
mlfn 32.5 2.8 softmax (256, 128) random_flip, random_crop euclidean 90.1 (74.3) 81.1 (63.2) 66.4 (37.2)
hacnn* 4.5 0.5 softmax (160, 64) random_flip, random_crop euclidean 90.9 (75.6) 80.1 (63.2) 64.7 (37.2)
mobilenetv2_x1_0 2.2 0.2 softmax (256, 128) random_flip, random_crop euclidean 85.6 (67.3) 74.2 (54.7) 57.4 (29.3)
mobilenetv2_x1_4 4.3 0.4 softmax (256, 128) random_flip, random_crop euclidean 87.0 (68.5) 76.2 (55.8) 60.1 (31.5)
osnet_x1_0 2.2 0.98 softmax (256, 128) random_flip euclidean 94.2 (82.6) 87.0 (70.2) 74.9 (43.8)
osnet_x0_75 1.3 0.57 softmax (256, 128) random_flip euclidean 93.7 (81.2) 85.8 (69.8) 72.8 (41.4)
osnet_x0_5 0.6 0.27 softmax (256, 128) random_flip euclidean 92.5 (79.8) 85.1 (67.4) 69.7 (37.5)
osnet_x0_25 0.2 0.08 softmax (256, 128) random_flip euclidean 91.2 (75.0) 82.0 (61.4) 61.4 (29.5)

Cross-domain ReID

Market1501 -> DukeMTMC-reID

Model # Param (10^6) GFLOPs Loss Input Transforms Distance Rank-1 Rank-5 Rank-10 mAP Download
osnet_ibn_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter euclidean 48.5 62.3 67.4 26.7 model
osnet_ain_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter cosine 52.4 66.1 71.2 30.5 model

DukeMTMC-reID -> Market1501

Model # Param (10^6) GFLOPs Loss Input Transforms Distance Rank-1 Rank-5 Rank-10 mAP Download
osnet_ibn_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter euclidean 57.7 73.7 80.0 26.1 model
osnet_ain_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter cosine 61.0 77.0 82.5 30.6 model

MSMT17 (combineall=True) -> Market1501 & DukeMTMC-reID

Model # Param (10^6) GFLOPs Loss Input Transforms Distance msmt17 -> market1501 msmt17 -> dukemtmcreid Download
resnet50 23.5 2.7 softmax (256, 128) random_flip, color_jitter euclidean 46.3 (22.8) 52.3 (32.1) model
osnet_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter euclidean 66.6 (37.5) 66.0 (45.3) model
osnet_x0_75 1.3 0.57 softmax (256, 128) random_flip, color_jitter euclidean 63.6 (35.5) 65.3 (44.5) model
osnet_x0_5 0.6 0.27 softmax (256, 128) random_flip, color_jitter euclidean 64.3 (34.9) 65.2 (43.3) model
osnet_x0_25 0.2 0.08 softmax (256, 128) random_flip, color_jitter euclidean 59.9 (31.0) 61.5 (39.6) model
osnet_ibn_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter euclidean 66.5 (37.2) 67.4 (45.6) model
osnet_ain_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter cosine 70.1 (43.3) 71.1 (52.7) model

Multi-source domain generalization

The models below are trained using multiple source datasets, as described in Zhou et al. TPAMI'21.

Regarding the abbreviations, MS is MSMT17; M is Market1501; D is DukeMTMC-reID; and C is CUHK03.

All models were trained with im_osnet_ain_x1_0_softmax_256x128_amsgrad_cosine.yaml and max_epoch=50.

Model # Param (10^6) GFLOPs Loss Input Transforms Distance MS+D+C->M MS+M+C->D MS+D+M->C D+M+C->MS
osnet_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter cosine 72.5 (44.2) 65.2 (47.0) 23.9 (23.3) 33.2 (12.6)
osnet_ibn_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter cosine 73.0 (44.9) 64.6 (45.7) 25.7 (25.4) 39.8 (16.2)
osnet_ain_x1_0 2.2 0.98 softmax (256, 128) random_flip, color_jitter cosine 73.3 (45.8) 65.6 (47.2) 27.4 (27.1) 40.2 (16.2)