zehuichen123/ddod

[ACM MM 2021, TMM 2023] Disentangle your Dense Object Detector

ddod's Issues

标签分配系数

大佬你好，想请教一下，论文中的table 5里面的reg和cls的alph，与成本矩阵中的alph是什么关系？不是alph与1-alph的关系吗。感谢

Batch size for COCO

For WIDER FACE, the paper describes the batch size.
"The model is trained with a batch size of 24 on 6 Titan V100s."

For COCO, is the total batch size 32?
tools/dist_train.sh coco_cfg/ddod_r50_1x.py 8
samples_per_gpu=4,

Can we achieve similar AP with a total batch size of 16 and lr=0.01?

羡慕啊，除了羡慕我还能说什么呢

柠檬，DDDD

如果不使用形变卷积，性能是否会下降，会下降多少？

如题

About Widerface

How is anchor set when ddod is applied to widerface?(including base_sizes, ratios, scales_per_octave, octave_base_scale)

GradCAM

The heatmap in fig.2. is so cool!

How to implement GradCAM in your detector?

在WiderFace上训练的mAP为0

Environment info：
sys.platform: linux
Python: 3.7.13 (default, Oct 18 2022, 18:57:03) [GCC 11.2.0]
CUDA available: True
GPU 0,1: NVIDIA A100 80GB PCIe
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 11.1, V11.1.74
GCC: gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0
PyTorch: 1.10.0
PyTorch compiling details: PyTorch built with:

• GCC 7.3
• C++ Version: 201402
• Intel(R) oneAPI Math Kernel Library Version 2021.4-Product Build 20210904 for Intel(R) 64 architecture applications
• Intel(R) MKL-DNN v2.2.3 (Git Hash 7336ca9f055cf1bfa13efb658fe15dc9b41f0740)
• OpenMP 201511 (a.k.a. OpenMP 4.5)
• LAPACK is enabled (usually provided by MKL)
• NNPACK is enabled
• CPU capability usage: AVX512
• CUDA Runtime 11.3
  TorchVision: 0.11.0
  OpenCV: 4.6.0
  MMCV: 1.6.2
  MMCV Compiler: GCC 9.3
  MMCV CUDA Compiler: 11.3
  MMDetection: 2.25.3+unknown

---

您好，我在mmdetection上根据您论文里Implementation Details的介绍（关于训练widerface），对ddod的优化器部分进行了以下改动：
optimizer = dict(type='SGD', lr=0.0075, momentum=0.9, weight_decay=0.0005)
optimizer_config = dict(grad_clip=None)
lr_config = dict(
policy='CosineRestart',
warmup='linear',
warmup_iters=500,
warmup_ratio=0.1,
periods=[
30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30,
30, 30
],
restart_weights=[
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1
],
min_lr_ratio=0.01)
runner = dict(type='EpochBasedRunner', max_epochs=600)
此外，我也将reg loss改为了DIoU loss，但是每个epoch报告的mAP均为0。请问是什么情况呢？

请问可以放出用于训练WIDER FACE的config文件吗？

r50换成r34检测不出目标

请问将backbone由r50换成r34检测不出目标是什么原因呢？

The meaning of "X" in "DDOD-X" at the last row of Table 8 in your paper?

Thank you for your great work! As the title says, I have a little question about "X".

About WiderFace.

When ddod is applied to wideface, does the testing process adopt the method in tinaface?
In addition, could you release the code to apply ddod to widerface?

FCOS version of DDOD

Hi, thanks for your great work!.

It seems that there is no FCOS version of DDOD (Use anchor point instead of anchor) implemented as indicated in the Table 3 of your paper, could you also share that part of code and config for reference?

Thanks!

COCO pre-trained models

Hi! Awesome work!
Do you have any plans to upload COCO weights?

Especially,

• ATSS(IoU) R-50 1x
• DDOD R-50 1x
• DDOD R2-101-DCN 2x
• DDOD-X R2-101-DCN 2x

It would be great if you also could provide configs for DDOD-X because the arXiv v1 seems to lack the explanation of DDOD-X and the details of multi-scale testing and stronger data augmentation.

Missing retina_assigner?

Hi, thanks for your great work. But when I run this code, it threw an error with AttributeError: 'ConfigDict' object has no attribute 'retina_assigner'. I guess self.train_cfg.retina_assigner in atss_r50_1x.py is missing. Can you please provide it?