PyTorch_YOLOF

A PyTorch version of You Only Look at One-level Feature object detector.

The input image must be resized to have their shorter side being 800 and their longer side less or equal to 1333.

During reproducing the YOLOF, I found many tricks used in YOLOF but the baseline RetinaNet dosen't use those tricks. For example, YOLOF takes advantage of RandomShift, CTR_CLAMP, large learning rate, big batchsize(like 64), negative prediction threshold. Is it really fair that YOLOF use these tricks to compare with RetinaNet?

In a other word, whether the YOLOF can still work without those tricks?

Requirements

• We recommend you to use Anaconda to create a conda environment:

conda create -n yolof python=3.6

• Then, activate the environment:

conda activate yolof

• Requirements:

pip install -r requirements.txt 

PyTorch >= 1.1.0 and Torchvision >= 0.3.0

Visualize positive sample

You can run following command to visualize positiva sample:

python train.py \
        -d voc \
        --batch_size 2 \
        --root path/to/your/dataset \
        --vis_targets

My Ablation Studies

image mask

• Backbone: ResNet-50
• image size: shorter size = 800, longer size <= 1333
• Batch size: 16
• lr: 0.01
• lr of backbone: 0.01
• SGD with momentum 0.9 and weight decay 1e-4
• Matcher: IoU Top4 (Different from the official matcher that uses top4 of L1 distance.)
• epoch: 12 (1x schedule)
• lr decay: 8, 11
• augmentation: RandomFlip

We ignore the loss of samples who are not in image.

L1 Top4

• Backbone: ResNet-50
• image size: shorter size = 800, longer size <= 1333
• Batch size: 16
• lr: 0.01
• lr of backbone: 0.01
• SGD with momentum 0.9 and weight decay 1e-4
• epoch: 12 (1x schedule)
• lr decay: 8, 11
• augmentation: RandomFlip
• with image mask

IoU topk: We choose the topK of IoU between anchor boxes and labels as the positive samples.

L1 topk: We choose the topK of L1 distance between anchor boxes and labels as the positive samples.

RandomShift Augmentation

• Backbone: ResNet-50
• image size: shorter size = 800, longer size <= 1333
• Batch size: 16
• lr: 0.01
• lr of backbone: 0.01
• SGD with momentum 0.9 and weight decay 1e-4
• Matcher: L1 Top4
• epoch: 12 (1x schedule)
• lr decay: 8, 11
• augmentation: RandomFlip
• with image mask

YOLOF takes advantage of RandomShift augmentation which is not used in RetinaNet.

Fix a bug in dataloader

• Backbone: ResNet-50
• image size: shorter size = 800, longer size <= 1333
• Batch size: 16
• lr: 0.01
• lr of backbone: 0.01
• SGD with momentum 0.9 and weight decay 1e-4
• Matcher: L1 Top4
• epoch: 12 (1x schedule)
• lr decay: 8, 11
• augmentation: RandomFlip + RandomShift
• with image mask

I fixed a bug in dataloader. Specifically, I set the shuffle in dataloader as False ...

Ignore samples

• Backbone: ResNet-50
• image size: shorter size = 800, longer size <= 1333
• Batch size: 16
• lr: 0.01
• lr of backbone: 0.01
• SGD with momentum 0.9 and weight decay 1e-4
• Matcher: L1 Top4
• epoch: 12 (1x schedule)
• lr decay: 8, 11
• augmentation: RandomFlip + RandomShift
• with image mask

We ignore those negative samples whose IoU with labels are higher the ignore threshold (igt).

Decode boxes

• Backbone: ResNet-50
• image size: shorter size = 800, longer size <= 1333
• Batch size: 16
• lr: 0.01
• lr of backbone: 0.01
• SGD with momentum 0.9 and weight decay 1e-4
• Matcher: L1 Top4
• epoch: 12 (1x schedule)
• lr decay: 8, 11
• augmentation: RandomFlip + RandomShift
• with image mask

Method-1: ctr_x = x_anchor + t_x, ctr_y = y_anchor + t_y

Method-2: ctr_x = x_anchor + t_x * w_anchor, ctr_y = y_anchor + t_y * h_anchor

The Method-2 is following the operation used in YOLOF.

Train

sh train.sh

You can change the configurations of train.sh.

If you just want to check which anchor box is assigned to the positive sample, you can run:

python train.py --cuda -d voc --batch_size 8 --vis_targets

According to your own situation, you can make necessary adjustments to the above run commands

Test

python test.py -d [select a dataset: voc or coco] \
               --cuda \
               -v [select a model] \
               --weight [ Please input the path to model dir. ] \
               --img_size 800 \
               --root path/to/dataset/ \
               --show

You can run the above command to visualize the detection results on the dataset.