Skip to content

ReID Models

Use this page to compare ReID backbones by compute, parameter count, embedding size, and available same-domain evaluation results.

Model Reference

Same-domain scores are reported as rank-1 (mAP). Rows with - do not yet have a documented same-domain result in this table. GFLOPs are input-size dependent and should be read together with the Input column; rows are not normalized to one common crop.

Model GFLOPs Input Params Embedding Market1501 CUHK03-L CUHK03-D DukeMTMC-reID MSMT17
osnet_x0_25 0.08 256x128 0.59M 512 91.2 (75.0) - - 82.0 (61.4) 61.4 (29.5)
osnet_x0_5 0.27 256x128 1.02M 512 92.5 (79.8) - - 85.1 (67.4) 69.7 (37.5)
osnet_x0_75 0.57 256x128 1.67M 512 93.7 (81.2) - - 85.8 (69.8) 72.8 (41.4)
mobilenetv4_conv_small 0.72 384x128 6.14M 1536 - - - - -
osnet_x1_0 0.98 256x128 2.56M 512 94.2 (82.6) - - 87.0 (70.2) 74.9 (43.8)
mobilenetv4_conv_medium 2.02 384x128 12.12M 1536 - - - - -
mobilenetv4_conv_large 4.68 384x128 35.02M 1536 - - - - -
lmbn_n 4.87 384x128 9.15M 3584 96.3 (91.5) 87.2 (85.1) 84.9 (82.4) - -
lmbn_ain_n 4.87 384x128 9.15M 3584 - - - - -
csl_tinyvit_7m 5.18 384x128 9.79M 1536 - - - - -
csl_tinyvit_11m 8.29 384x128 15.44M 1536 - - - - -
csl_tinyvit_23m 15.15 384x128 25.71M 1536 - - - - -

The linked OSNet scores use the published same-domain softmax setup with 256x128 input, random flip augmentation, and Euclidean distance. The lmbn_n scores are the LightMBN paper row reported as Ours LightMBN with an OSNet backbone. MobileNetV4, LMBN, and CSL-TinyViT rows use the 384x128 BoxMOT ReID crop used by their training configs.