Can you guys reproduce the performance??

[lhtht77]

I trained yolov9-m model with coco dataset.
The detector's performance stuck on about 36AP.

The official code goes up to about 51 AP.

The difference is so big and the root cause is not clear.

Anybody overcome this issue before?

[sam31046]

You could try augmentation or remove some. See this issue #197

[BitCode-in]

Hi everyone,

I had the exact same issue with low metrics on a custom dataset (small objects on complex backgrounds) and wanted to share what actually fixed it for me.

@lhtht77 If your repo doesn't support color augs from the yaml, you have to write a custom transform using torchvision. Here is a quick example for brightness that you can add to your data augmentation script:

import torchvision.transforms.functional as TF
import numpy as np
import torch

class RandomBrightness:
    def __init__(self, prob=0.5, factor_range=[0.8, 1.2]):
        self.prob = prob
        self.factor_range = factor_range

    def __call__(self, image, boxes):
        if torch.rand(1) < self.prob:
            factor = np.random.uniform(self.factor_range[0], self.factor_range[1])
            image = TF.adjust_brightness(image, factor)
        return image, boxes

Just make sure your yaml parser unpacks the dict parameters correctly so prob doesn't accidentally get passed as a config object.

To save you all some time, here is my complete train.yaml. Switching from SGD to AdamW (with lr 0.001) fixed the convergence speed for me.

But I also want to highly recommend looking at the scheduler block. Switching to CosineAnnealingLR instead of Linear was another massive leap for my metrics.

task: train

defaults:
  - validation: ../validation

epoch: 500

data:
  batch_size: 16
  equivalent_batch_size: 64
  image_size: ${image_size} 
  cpu_num: ${cpu_num}
  shuffle: True
  pin_memory: True
  
  data_augment:
    Mosaic: 1.0
    MixUp: 0.1
    HorizontalFlip: 0.5
    RandomCrop: 0.0
    VerticalFlip: 0.0
    
    RandomBrightness:
      prob: 0.5
      factor_range: [0.6, 1.4]
    RandomContrast:
      prob: 0.5
      factor_range: [0.6, 1.4]
    RandomSaturation:
      prob: 0.5
      factor_range: [0.6, 1.4]
    
    RemoveOutliers: 1e-8

optimizer:
  type: AdamW
  args:
    lr: 0.001
    weight_decay: 0.05
    betas: [0.9, 0.999]

loss:
  objective:
    BCELoss: 0.5
    BoxLoss: 7.5
    DFLoss: 1.5
  aux: 0.25
  matcher:
    iou: CIoU
    topk: 10
    factor:
      iou: 6.0
      cls: 0.5

scheduler:
  type: CosineAnnealingLR
  warmup:
    epochs: 3.0
  args:
    T_max: ${task.epoch}
    eta_min: 0.0001

ema:
  enable: true
  decay: 0.9999

Hope this helps you guys get those metrics up.