chore(bevfusion): update parameters for improved bevfusion-cl training #88

SamratThapa120 · 2025-08-26T05:44:03Z

This pull request introduces significant improvements and updates to the BEVFusion-CL base and offline configurations, focusing on data preprocessing, augmentation, model configuration, and documentation. The changes enhance training stability, improve augmentation consistency, and update documentation to reflect the latest evaluation results.

Key changes include:

Data Augmentation and Preprocessing

Refactored the sample_augmentation method in transforms_3d.py to handle scalar resize_lim values, ensuring consistent resizing behavior and more robust augmentation during training and testing. Rotation now uses bicubic resampling to reduce artifacts. (projects/BEVFusion/bevfusion/transforms_3d.py) [1] [2] [3]
Updated the ImageAug3D pipeline in both training and testing to use a maximum resize limit scalar resize_lim=0.02, allowing training with images with various aspect ratios. (projects/BEVFusion/configs/t4dataset/BEVFusion-CL-offline/bevfusion_camera_lidar_offline_voxel_second_secfpn_4xb8_base.py, projects/BEVFusion/configs/t4dataset/BEVFusion-CL/bevfusion_camera_lidar_voxel_second_secfpn_4xb8_base.py) [1] [2] [3] [4]

Model and Config Updates

Major overhaul of the BEVFusion-CL-offline config: increased train_gpu_size, adjusted image_size, feature_size, dbound, and other model parameters for better performance and scalability; added filter_cfg to filter frames with missing images; and enabled automatic learning rate scaling. (projects/BEVFusion/configs/t4dataset/BEVFusion-CL-offline/bevfusion_camera_lidar_offline_voxel_second_secfpn_4xb8_base.py) [1] [2] [3] [4] [5] [6] [7] [8] [9]
Updated the BEVFusion-CL base config with new image_size, feature_size, and dbound values, and enabled automatic learning rate scaling. (projects/BEVFusion/configs/t4dataset/BEVFusion-CL/bevfusion_camera_lidar_voxel_second_secfpn_4xb8_base.py) [1] [2] [3]

Documentation and Evaluation Results

Added a new documentation page summarizing the deployed BEVFusion-CL-offline base/2.X model, including training and evaluation details, metrics, and links to resources. (projects/BEVFusion/docs/BEVFusion-CL-offline/v2/base.md)
Updated the BEVFusion-CL base documentation to include results for a new evaluation split (C). (projects/BEVFusion/docs/BEVFusion-CL/v2/base.md)

These changes collectively improve the robustness, reproducibility, and clarity of the BEVFusion-CL and BEVFusion-CL-offline pipelines, and provide up-to-date documentation for users and collaborators.

Improvement in bevfusion-CL base/2.0.0 before and after the changes

Eval range: 120m	mAP	car	truck	bus	bicycle	pedestrian
BEVFusion-CL base/2.0.0 (B)	75.03	79.62	61.20	86.67	69.99	77.62
BEVFusion-CL base/2.0.0 (C)	76.3	80.50	61.90	85.90	74.70	78.70
BEVFusion-CL-offline base/2.0.0 (C)	77.8	87.30	61.60	85.90	73.20	80.90

BEVFusion-CL base/2.0.0 (B): Without intensity and training pedestrians without pooling
BEVFusion-CL base/2.0.0 (C): Same as BEVFusion-CL base/2.0.0 (B) with improved image ROI cropping, and augmentation parameter fixes.

Signed-off-by: Samrat Thapa <[email protected]>

KSeangTan

LGTM overall, just need to tidy up documentation a little bit

KSeangTan · 2025-09-10T05:18:01Z

projects/BEVFusion/bevfusion/transforms_3d.py

        if flip:
            img = img.transpose(method=Image.FLIP_LEFT_RIGHT)
-        img = img.rotate(rotate)
+        img = img.rotate(rotate, resample=Image.BICUBIC)  # Default rotation introduces artifacts.


Would be nice if you have examples showing artifacts with the default rotation

KSeangTan · 2025-09-10T05:30:30Z

...dataset/BEVFusion-CL-offline/bevfusion_camera_lidar_offline_voxel_second_secfpn_4xb8_base.py

        zbound=[-10.0, 10.0, 20.0],
-        # dbound=[1.0, 60.0, 0.5],
-        dbound=[1.0, 166.2, 1.4],
+        dbound=[1.0, 134, 1.4],


Any reason we change it to 134, I am thinking we should make the depth and bin size even smaller, and make sure it's evenly divided by the bin size?

bin size: 1.4 could be a little big too large

KSeangTan · 2025-09-10T05:32:55Z

...dataset/BEVFusion-CL-offline/bevfusion_camera_lidar_offline_voxel_second_secfpn_4xb8_base.py

-#   - `base_batch_size` = (8 GPUs) x (4 samples per GPU).
-# auto_scale_lr = dict(enable=False, base_batch_size=32)
-auto_scale_lr = dict(enable=False, base_batch_size=train_gpu_size * train_batch_size)
+auto_scale_lr = dict(enable=True, base_batch_size=4)


Keep the comment, and any reason we set it to True? Does it show any significant improvement/stability for training?

KSeangTan · 2025-09-10T05:34:17Z

...dataset/BEVFusion-CL-offline/bevfusion_camera_lidar_offline_voxel_second_secfpn_4xb8_base.py

 if train_gpu_size > 1:
    sync_bn = "torch"
-
-randomness = dict(seed=0, diff_rank_seed=False, deterministic=True)


Any reason we delete it? I believe we need to keep it for reproducibility

KSeangTan · 2025-09-10T05:35:20Z

...usion/configs/t4dataset/BEVFusion-CL/bevfusion_camera_lidar_voxel_second_secfpn_4xb8_base.py

 #   - `base_batch_size` = (8 GPUs) x (4 samples per GPU).
-# auto_scale_lr = dict(enable=False, base_batch_size=32)
-auto_scale_lr = dict(enable=False, base_batch_size=train_gpu_size * train_batch_size)
+auto_scale_lr = dict(enable=True, base_batch_size=32)


base_batch_size should be batch size per gpu, which should be train_batch_size according to here
https://github.com/open-mmlab/mmengine/blob/main/mmengine/_strategy/base.py#L696

KSeangTan · 2025-09-10T05:47:27Z

projects/BEVFusion/docs/BEVFusion-CL/v2/base.md


+- BEVFusion-CL base/2.0.0 (A): Without intensity and training pedestrians with pooling pedestrians
+- BEVFusion-CL base/2.0.0 (B): Same as `BEVFusion-CL base/2.0.0 (A)` without pooling pedestrians
+- BEVFusion-CL base/2.0.0 (C): Same as `BEVFusion-CL base/2.0.0 (B)` with improved image ROI cropping, and augmentation parameter fixes.


I suppose you meant
BEVFusion-CL base/2.0.0 (A) is without pooling pedestrians, and BEVFusion-CL base/2.0.0 (B) with pooling pedestrians? Otherwise, the performance in pedestrians doesn't make sense to me

KSeangTan · 2025-09-10T05:50:11Z

...usion/configs/t4dataset/BEVFusion-CL/bevfusion_camera_lidar_voxel_second_secfpn_4xb8_base.py

        zbound=[-10.0, 10.0, 20.0],
-        # dbound=[1.0, 60.0, 0.5],
-        dbound=[1.0, 166.2, 1.4],
+        dbound=[1.0, 134, 1.4],


SamratThapa120 and others added 3 commits August 26, 2025 14:37

updated roi params for t4dataset

2df7b08

Signed-off-by: Samrat Thapa <[email protected]>

remove unused param

354ff1b

ci(pre-commit): autofix

dc1d30f

SamratThapa120 changed the title ~~chore(bevfusion): update image rois appropriately for t4datasets~~ chore(bevfusion): update parameters for improved bevfusion-cl training Aug 26, 2025

SamratThapa120 and others added 3 commits August 26, 2025 16:08

added bevfusion l

09f35d5

Signed-off-by: Samrat Thapa <[email protected]>

update params

50f35a8

updated docs

a55a907

SamratThapa120 requested a review from KSeangTan September 8, 2025 09:04

pre-commit-ci bot and others added 2 commits September 8, 2025 09:05

ci(pre-commit): autofix

7006d5a

Merge branch 'main' into chore/bevfusion/update_rois

8e74712

SamratThapa120 marked this pull request as ready for review September 8, 2025 09:06

fixed lr scaler

a038999

SamratThapa120 mentioned this pull request Sep 9, 2025

feat(fix): fix bevfusion deployment script #98

Open

KSeangTan requested changes Sep 10, 2025

View reviewed changes

KSeangTan assigned SamratThapa120 Sep 10, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

chore(bevfusion): update parameters for improved bevfusion-cl training #88

chore(bevfusion): update parameters for improved bevfusion-cl training #88

Uh oh!

SamratThapa120 commented Aug 26, 2025 •

edited

Loading

Uh oh!

KSeangTan left a comment

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

KSeangTan Sep 10, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

chore(bevfusion): update parameters for improved bevfusion-cl training #88

Are you sure you want to change the base?

chore(bevfusion): update parameters for improved bevfusion-cl training #88

Uh oh!

Conversation

SamratThapa120 commented Aug 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Data Augmentation and Preprocessing

Model and Config Updates

Documentation and Evaluation Results

Uh oh!

KSeangTan left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

SamratThapa120 commented Aug 26, 2025 •

edited

Loading