You are seeing it wrong. Please gaze in a right direction π
Description: Person Re-ID becomes a challenging task when there are lot of disturbances in the image such as occlusion, person-box aspect ratios, crop scales, and training the re-id models with present approaches of supervised classification and triplet matching when the number of classes are very high. Training the models on the public datasets like Duke-MTMC and Market1501 does not perform well on the issues mentioned as they are prominent in the real-world scenarios. I, therefore to tackle these problems, designed a Re-ID model which can take into consideration the box scale and alignment of a person crop for the best retrieval accuracy. I tested it on many datasets and I am mentioning the results on Market1501 using the proposed algorithm. Also, there was a significant gain observed when it was tested on the private companyβs datasets.
Model | Data | Rank@1 | Rank@5 | Rank@10 | mAP |
---|---|---|---|---|---|
Baseline | Market1501 | 94.63 | 98.25 | 99.05 | 85.24 |
Baeline + Alignment + MultiScale | Market1501 | 96.59 | 98.84 | 99.44 | 91.24 |