Person re-identification is the process of associating images of the same person which is used for image retrieval, given a query or person-of-interest. In this project, we used ResNet50 and DenseNet121 deep learning models to achieve the preson re ID. We used a combination of Triplet loss and Softmax loss to improve the accuracy. We used Random erasing as a data augmentation technique to enhance the data set.
Model: ResNet50 Loss: Softmaxloss
On market1501 Test-set: Rank1:85.74 Rank5: 95.19 Rank10: 96.94 mAP: 67.3
======================================================================================================================
Model: ResNet50 Loss:Softmax Loss Added: Random Erasing
On market1501 Test-set: Rank1:88.15 Rank5:95.8432 Rank10:97.1793 mAP:72.3722
======================================================================================================================
Model: ResNet50 Loss:Triplet Loss
On Val-set provided by TA: top1: 89.33, top5: 94.67, top10: 97.33, mAP: 74.59 On market1501 Test-set: Rank1:89 Rank5:93.4 Rank10:97 mAP:72.33
======================================================================================================================
Model: DenseNet121 Loss:Triplet Loss (No Random erasing )
Val-Set provided by TA: top1: 86.67, top5: 95.33, top10: 99.33, mAP: 76.28
======================================================================================================================
Model: DenseNet121 Loss: Triplet Loss
On Val-set provided by TA: top1: 92.00, top5: 98.00, top10: 99.33, mAP: 78.68 On market1501 Test-set: top1: 89.8, top5: 96.30, top10: 98.33, mAP: 75.2
======================================================================================================================
Model: MidResNet50 Loss: Triplet Loss
On Val-set provided by TA: top1: 88.00, top5: 94.567, top10: 94.67, mAP: 73.87
======================================================================================================================
DenseNet121 gave best result, so we added softmax loss and ran all the experiments on denset121 only.
| (triplet_loss/softmax_loss) weightage | top1 (val/test) % | top5 (val/test) % | top10 (val/test) % | mAP (val/test) % |
|---|---|---|---|---|
| 0.9/0.1 | 90/90.5 | 95.33/96.7 | 96.67/98.01 | 78.53/77.27 |
| 0.85/0.15 | 94/89.9 | 97.33/96 | 99.33/97.6 | 79.46/76.82 |
| 0.8/0.2 | 91.33/90.1 | 98/96.46 | 97.67/97.86 | 78.05/76.878 |
| 0.2/0.8 | 91.33/91.4 | 97.33/97 | 98.67/98.18 | 79.09/78.8 |
| 0.1/0.9 | 91.33/92 | 96.67/97.2 | 98/98.2 | 80.11/78.25 |
| 0.1/0.9 - Features Normalize | 93.33/91.7 | 97.33/96.88 | 98.67/98 | 81.74/78.97 |
A .pth file has been uploaded at https://drive.google.com/open?id=1AT-x3HCVWAuZNEqSFvqbpWjnF2mzBYx7