Discrepancy in Calculated Parameters (≈45M) vs. Paper Reported Parameters (30.5M) for Strip R-CNN-S

[cs-an]

Thank you for open-sourcing the Strip R-CNN-S model and related tools.

When trying to reproduce the model complexity metrics (parameters and FLOPs) for the Strip R-CNN-S model in your paper, I observed a significant discrepancy in the parameter count compared to the value reported in Table 3 and Table 6 when using the provided analysis_tools/get_flops.py script.

---

Steps to Reproduce:

1. Calculate FLOPs/Params for the HRSC configuration (strip_rcnn_s_fpn_3x_hrsc_le90.py) with input shape 800x800:

   python tools/analysis_tools/get_flops.py configs/strip_rcnn/strip_rcnn_s_fpn_3x_hrsc_le90.py --shape 800

   Observed Output:

   Input shape: (3, 800, 800)
   Flops: 157.13 GFLOPs
   Params: 45.12 M
   

2. Calculate FLOPs/Params for the DOTA configuration (e.g., strip_rcnn_s_fpn_1x_dota_le90.py or similar standard S config) with input shape 1024x1024:

   python tools/analysis_tools/get_flops.py configs/strip_rcnn/strip_rcnn_s_fpn_1x_dota_le90.py --shape 1024

   Observed Output:

   Input shape: (3, 1024, 1024)
   Flops: 218.62 GFLOPs
   Params: 45.14 M
   

---

Discrepancy and Question:

The calculated parameter count (around 45M) in my environment is significantly higher than the reported 30.5M in the paper (Table 3 & 6) for the Strip R-CNN-S model. This discrepancy appears consistently across both the HRSC and DOTA configurations.

Could you please clarify if this parameter count difference is expected due to code updates/refactoring after the paper was published, or if there might be a specific configuration or setup step I'm missing to match the reported 30.5M parameters?

---

Thank you for your time and the great work on this project!

[YXB-NKU]

The 45M you calculated should be due to the incorrect setting of the parameters in the Strip head section. In the code we provided, the default parameters for the Strip head, the angle head and the classification head do not share fully connected layers.

[cs-an]

Hello,

Based on the provided information, I modified num_reg_theta_fcs in StripHead from 2 to 0. Subsequently, I executed the following command:

python tools/analysis_tools/get_flops.py work_dirs/strip_rcnn/strip_rcnn_s_fpn_3x_hrsc_le90/strip_rcnn_s_fpn_3x_hrsc_le90.py --shape 800

The output I received is:

==============================
Input shape: (3, 800, 800)
Flops: 143.25 GFLOPs
Params: 31.24 M
==============================

This result still seems inconsistent with the values reported in the paper. I have two questions regarding this:

1、Could you please clarify the method used to calculate the model's FLOPs and parameter count as presented in the paper? Was the get_flops.py script used similarly? If possible, could you provide the specific details or configuration needed to replicate the reported results accurately?

2、Regarding the results tables in the paper: could you explain why the reported FLOPs/Params values appear to be the same for experiments on both the HRSC2016 dataset (using an 800x800 resolution) and the DOTAv1.0 dataset (using a 1024x1024 resolution), despite the difference in input shapes?

Any clarification you could provide on these points would be greatly appreciated.
Thank you for your time and assistance.

[cs-an]

Hi @YXB-NKU, sorry to bother you again, but I was just wondering if you might have had a chance to see my previous question above about FLOPs/Params? Any insights you could offer would be greatly appreciated when you have a moment. Thanks!

[YXB-NKU]

Hello , I'm busy on another project these days. For the FLops/Params computation on HRSC2016 and DOTAv1.0 dataset I keep the resolution of 1024x1024，so it's the same.

[cs-an]

Hello, thanks again for your great work on Strip-R-CNN!

Following your suggestion, I set num_reg_theta_fcs = 0 in the StripHead configuration. However, when I used tools/analysis_tools/get_flops.py to check the model size, the results were as follows:

==============================
Input shape: (3, 1024, 1024)
Flops: 204.73 GFLOPs
Params: 31.25 M
==============================

==============================
Input shape: (3, 800, 800)
Flops: 143.26 GFLOPs
Params: 31.25 M
==============================

I'm trying to accurately reproduce your setup, but these results are somewhat confusing to me. Thank you very much for your time and any clarification you can provide.

[cappuccino-wxs]

Hello, thanks again for your great work on Strip-R-CNN!

Following your suggestion, I set num_reg_theta_fcs = 0 in the StripHead configuration. However, when I used tools/analysis_tools/get_flops.py to check the model size, the results were as follows:

==============================
Input shape: (3, 1024, 1024)
Flops: 204.73 GFLOPs
Params: 31.25 M
==============================

==============================
Input shape: (3, 800, 800)
Flops: 143.26 GFLOPs
Params: 31.25 M
==============================

I'm trying to accurately reproduce your setup, but these results are somewhat confusing to me. Thank you very much for your time and any clarification you can provide.

When num_reg_theta_fcs is set to 0 in the StripHead configuration, the input dimension of the angle head becomes 12544 (from 256×7×7 feature maps). According to the pretrained model "strip_rcnn_s_dota.pth", this dimension should be 1024. After adjusting the input dimension to 1024, the model structure aligns with the paper, and Strip R-CNN S has a final parameter count of 31.17M.