More Evaluations on Re identification w . Market1501

Experiments

• SeRes18-IBN-BatchReNorm(Important!)-FocalLoss(Important!)+TripletPenalty+Center-Continual

  Metric	Acc@1	Acc@5	Acc@10	mAP	Size
  Value	0.8593	0.9501	0.9697	0.6493	44.246 MB

• SeRes18-IBN-BatchReNorm(Important!)-FocalLoss(Important!)+TripletPenalty+Center

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	0.8530	0.9483	0.9667	0.6409

• SeRes18-IBN-BatchReNorm(Important!)-PolyLoss(Important; epsilon=1.0)+TripletPenalty+Center-Continual(thres=0.3)

  Checkpoint

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	0.8473	0.9445	0.9685	0.6401

• SeRes18-IBN-BatchReNorm(Important!)-PolyLoss(Important; epsilon=1.0)+TripletPenalty+Center

  Checkpoint

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	0.8438	0.9424	0.9685	0.6365

• SeRes18-IBN-BatchReNorm(Important!)-PolyLoss(Important; epsilon=1.0)+SoftTriplet+Center (Pending Update)

  Checkpoint

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	-	-	-	-

• SERes18-IBN-BatchReNorm(Important!)-Softmax+TripletPenalty+Center (Pending Update)

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	-	-	-	-

• SeRes18-IBN-BatchReNorm(Important!)-FocalLoss(Important!)+SoftTriplet+Center

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	0.8189	0.9317	0.9507	0.6002

• SeRes18-IBN-BatchReNorm(Important!)-Softmax+SoftTriplet+Center-ContinualLearning

  Checkpoint

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	0.8138	0.9311	0.9531	0.5924

• SeRes18-IBN-BatchReNorm(Important!)-Softmax+SoftTriplet+Center

  Checkpoint

  Metric	Acc@1	Acc@5	Acc@10	mAP
  Value	0.8058	0.9258	0.9528	0.5848

Conclusion

The proposal of these bag of tricks and bag of freebies will boost re-identification accuracy in a small-batch class-imbalance dataset.