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 - - - - -
hi_afa 2.24† 384x128 12.76M† 8192 97.0 (91.8)† - - 91.7 (82.9)† 87.6 (71.9)†
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_7m_v20 3.29‡ 384x128 6.94M‡ 1152 95.6 (90.0)‡ - - - -
csl_tinyvit_11m 8.29 384x128 15.46M 1536 - - - - -
csl_tinyvit_11m_v20 5.85 384x128 13.51M 1536 95.9 (91.0)§ - - - -
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, Hi-AFA, and CSL-TinyViT rows use the 384x128 BoxMOT ReID crop used by their training configs.

† Hi-AFA complexity and scores are reported by the Hi-AFA paper; they are not a BoxMOT checkpoint result. The current implementation has 11.78M parameters with Market1501's 751 classifiers. It emits the 8192-D balanced_unique_raw_pooled_streams descriptor: 16 unique 512-D streams are normalized independently before concatenation and final normalization, and the evaluation-only DropBlock duplicate of g4 is omitted. The paper does not publish code and leaves attention gamma, DropBlock geometry, label smoothing, its learning-rate schedule, and shared versus branch-specific center tables unspecified. The stabilized hi_afa_market1501 recipe uses zero-initialized, trainable LDAM residual gates and disables the ambiguous shared 22-stream center objective while recording BoxMOT's other explicit reproduction assumptions.

‡ csl_tinyvit_7m_v20 names the promoted RGB inference topology explicitly. Its 6,937,893 parameters count a normal 751-class training construction without privileged anatomy; runtime deployment removes classifiers and folds BNNecks. The full V20 training recipe enables a 227,118-parameter, training-only anatomy teacher and therefore contains 7,165,011 parameters. The reported Market1501 result is 95.58 rank-1 / 89.96 mAP from that recipe's compact FP16 checkpoint reloaded for RGB inference.

csl_tinyvit_11m is the configurable generic 11M constructor. csl_tinyvit_11m_v20 fixes the promoted semantic-fine, depthwise-separable, scale-balanced topology under an explicit checkpoint-safe name.

§ The 95.90 rank-1 / 91.02 mAP Market1501 result was produced by the A11v8 multiscale EMA pose teacher, now captured by the csl_tinyvit_11m_v20_pose_teacher training recipe. Its privileged branch raises the 751-class training construction from 13,514,597 to 13,887,539 parameters, but is pruned for deployment; inference remains the 13.51M, 1536-D RGB V20 model. The RGB-only csl_tinyvit_11m recipe is a separate training policy.