/VehicleReID

VehicleReID

Experiment Results

# code: /home/lmb/proj/vehicle-reid
# experiment: /media/disk1/lmb/EXP
# resnet50_fc256 pretrain model: /media/disk1/lmb/EXP/exp-vehicle-reid/train-ve1m+veid+veri/model_59.pt

Train in VehicleID (13.164k)

# /home/lmb/proj/vehicle-reid/train_id.sh
python train_xent_tri.py \
-s vehicleid \  # 训练集名称
-t vehicleid \  # 测试集名称
--height 128 \
--width 256 \
--optim amsgrad \
--lr 0.0003 \
--max-epoch 60 \
--stepsize 20 40 \
--train-batch-size 512 \
--test-batch-size 100 \
-a resnet50 \  # resnet50模型的特征维度2048
--save-dir /media/disk1/lmb/log/exp \
--gpu-devices "0, 1, 2, 3" \
--use-avai-gpus \
--test-size 800 \  # 测试集大小
--workers 4

# /home/lmb/proj/vehicle-reid/test_id.sh
python train_xent_tri.py \
-s vehicleid \  # 训练集名称
-t vehicleid \  # 测试集名称
--height 128 \
--width 256 \
--test-size 800 \
--train-batch-size 1 \
--test-batch-size 512 \
--evaluate \ 
-a resnet50 \ # resnet50模型的特征维度2048
--load-weights /media/disk1/lmb/EXP/exp-vehicle-reid/train-1m/model_best.pth \
--save-dir /media/disk1/lmb/EXP/exp-vehicle-reid/test-1m-id \
--root /media/disk1/lmb/DATASET/ \
--gpu-devices "0, 1, 2, 3" \
--use-avai-gpus
train (size) test (size) h w mAP Rank-1 Rank-5 Rank-10 Rank-20
VehicleID (13k) VehicleID (0.8k) 128 256 81.8 74.6 91.1 95.0 97.2
VehicleID (13k) VehicleID (1.6k) 128 256 77.5 70.3 86.3 91.3 94.6
VehicleID (13k) VehicleID (2.4k) 128 256 74.8 67.5 83.8 89.1 92.7
VehicleID (13k) VehicleID (3.2k) 128 256 72.2 64.9 80.6 86.4 91.2
VehicleID (13k) VehicleID (6.0k) 128 256 66.7 59.3 74.8 80.9 86.7
VehicleID (13k) VehicleID (13.164k) 128 256 61.6 54.2 69.6 75.2 80.9
train (size) test (size) h w mAP Rank-1 Rank-5 Rank-10 Rank-20
VehicleID (13k) Vehicle-1M (1k) 128 256 70.3 62.8 78.8 83.6 88.1
VehicleID (13k) Vehicle-1M (2k) 128 256 63.4 54.9 73.2 78.8 83.9
VehicleID (13k) Vehicle-1M (3k) 128 256 60.4 51.5 70.8 76.7 81.9
VehicleID (13k) Vehicle-1M (5.527k) 128 256 55.1 46.3 65.3 71.5 77.0

Train in Vehicle-1M (930k)

# /home/lmb/proj/vehicle-reid/train_1m.sh
python train_xent_tri.py \
-s vehicle1m \  # 训练集名称
-t vehicle1m \  # 测试集名称
--height 128 \
--width 256 \
--optim amsgrad \
--lr 0.0003 \
--max-epoch 60 \
--stepsize 20 40 \
--train-batch-size 256 \
--test-batch-size 100 \
-a resnet50 \ # resnet50模型的特征维度2048
--save-dir /media/disk1/lmb/EXP/exp-vehicle-reid/train-1m \
--gpu-devices "0, 1, 2, 3" \
--use-avai-gpus \
--test-size 1000 \  # 测试集大小
--workers 32 \
--root /media/disk1/lmb/DATASET/ \
--resume /media/disk1/lmb/EXP/exp-vehicle-reid/train-1m/model_9.pt

# /home/lmb/proj/vehicle-reid/test_1m.sh
python train_xent_tri.py \
-s vehicle1m \  # 训练集名称
-t vehicle1m \  # 测试集名称
--height 128 \
--width 256 \
--test-size 5527 \  # 测试集大小
--train-batch-size 1 \ # 不起作用
--test-batch-size 512 \
--evaluate \
-a resnet50 \
--load-weights /media/disk1/lmb/EXP/exp-vehicle-reid/train-1m/model_best.pth \
--save-dir /media/disk1/lmb/EXP/exp-vehicle-reid/test-1m \
--root /media/disk1/lmb/DATASET/ \
--gpu-devices "0, 1, 2, 3" \
--use-avai-gpus
train (size) test (size) h w mAP Rank-1 Rank-5 Rank-10 Rank-20
Vehicle-1M (930k) VehicleID (0.8k) 128 256 66.8 60.8 73.4 78.2 82.9
Vehicle-1M (930k) VehicleID (1.6k) 128 256 65.6 59.6 72.3 77.0 80.4
Vehicle-1M (930k) VehicleID (2.4k) 128 256 62.2 56.0 69.2 73.7 77.9
Vehicle-1M (930k) VehicleID (3.2k) 128 256 59.4 53.3 66.0 70.2 74.7
Vehicle-1M (930k) VehicleID (6.0k) 128 256 55.7 49.5 62.4 66.5 70.6
Vehicle-1M (930k) VehicleID (13.164k) 128 256 51.0 44.7 58.1 62.7 66.9
train (size) test (size) h w mAP Rank-1 Rank-5 Rank-10 Rank-20
Vehicle-1M (930k) Vehicle-1M (1k) 128 256 96.7 95.0 98.8 99.1 99.3
Vehicle-1M (930k) Vehicle-1M (2k) 128 256 92.6 90.7 94.9 95.6 96.3
Vehicle-1M (930k) Vehicle-1M (3k) 128 256 92.1 89.2 95.7 96.8 97.6
Vehicle-1M (930k) Vehicle-1M (5.527k) 128 256 90.6 86.7 95.4 96.9 97.7

Train in Vehicle-1M、VehicleID、VeRi (990k)

# /home/lmb/proj/vehicle-reid/train_ve1m+veid+veri.sh
python train_xent_tri.py \
-s vehicle1m vehicleid veri \  # 训练集名称（多个训练集）
-t vehicle1m vehicleid veri \  # 测试集名称（多个测试集）
--height 128 \
--width 256 \
--optim amsgrad \
--lr 0.0003 \
--max-epoch 60 \
--stepsize 20 40 \
--train-batch-size 256 \
--test-batch-size 100 \
-a resnet50_fc256 \  # 特征维度256
--save-dir /media/disk1/lmb/EXP/exp-vehicle-reid/train-ve1m+veid+veri \
--gpu-devices "0, 1, 2, 3" \
--use-avai-gpus \
--test-size 1000 \  # 测试集大小（多个测试集时，需要在代码里定义不同测试集的默认大小）
--workers 16 \
--root /media/disk1/lmb/DATASET/ \
--label-smooth

# /home/lmb/proj/vehicle-reid/test_ve1m+veid+veri.sh
python train_xent_tri.py \
-s vehicleid \
-t vehicleid \
--height 128 \
--width 256 \
--test-size 800 \
--train-batch-size 1 \
--test-batch-size 512 \
--evaluate \
-a resnet50_fc256 \  # 特征维度256
--load-weights /media/disk1/lmb/EXP/exp-vehicle-reid/train-ve1m+veid+veri/model_59.pt \
--save-dir /media/disk1/lmb/EXP/exp-vehicle-reid/test-ve1m+veid+veid \
--root /media/disk1/lmb/DATASET/ \
--gpu-devices "0, 1, 2, 3" \
--use-avai-gpus
train (size) test (size) h w mAP Rank-1 Rank-5 Rank-10 Rank-20
Vehicle-1M+VehicleID+VeRi (990k) VehicleID (0.8k) 128 256 85.7 80.2 93.2 96.8 98.0
Vehicle-1M+VehicleID+VeRi (990k) VehicleID (1.6k) 128 256 82.0 76.1 89.0 94.0 96.5
Vehicle-1M+VehicleID+VeRi (990k) VehicleID (2.4k) 128 256 78.4 72.4 85.6 91.2 95.2
Vehicle-1M+VehicleID+VeRi (990k) VehicleID (3.2k) 128 256 76.5 70.4 83.4 89.3 94.1
Vehicle-1M+VehicleID+VeRi (990k) VehicleID (6.0k) 128 256 73.9 68.3 80.0 84.8 89.9
Vehicle-1M+VehicleID+VeRi (990k) VehicleID (13.164k) 128 256 69.6 64.4 74.7 79.3 84.4
train (size) test (size) h w mAP Rank-1 Rank-5 Rank-10 Rank-20
Vehicle-1M+VehicleID+VeRi (990k) Vehicle-1M (1k) 128 256 96.1 93.9 98.8 99.2 99.5
Vehicle-1M+VehicleID+VeRi (990k) Vehicle-1M (2k) 128 256 93.7 91.9 95.8 96.3 96.6
Vehicle-1M+VehicleID+VeRi (990k) Vehicle-1M (3k) 128 256 93.4 91.0 96.2 97.0 97.7
Vehicle-1M+VehicleID+VeRi (990k) Vehicle-1M (5.527k) 128 256 91.5 87.9 95.9 96.9 97.6

Time

Query 15123 30539 45069 1 1 1 1
Gallery 1k 2k 3k 10^6 10^7 10^8 10^9
Distmat Times 0.8 2.98 6.32 0.04879 0.4879 4.879 48.79