Closed VictorWylde closed 3 months ago
Hi, can you show some images? Are they shown correctly during training in the tensorboard?
Sorry for the late reply. I would like to upload the training results, but there have been issues with GitHub's push. This is the opt.json for the inference results and training weights.
{
"data_path": "/workspace/IPC_Indoor_20231030_1530_P",
"log_dir": "/mnt/smb-149/02_Tasks/01_Depth_Estimation/Lite-Mono-main/tmp",
"model_name": "mytrain-010-lite-mono-8m_640_192-resnet50-separate_resnet",
"split": "odom",
"split_num": 3,
"model": "lite-mono-8m",
"weight_decay": 0.01,
"drop_path": 0.2,
"num_layers": 50,
"dataset": "lift_odom",
"png": false,
"height": 192,
"width": 640,
"disparity_smoothness": 0.001,
"scales": [
0,
1,
2
],
"min_depth": 0.1,
"max_depth": 100.0,
"use_stereo": false,
"frame_ids": [
0,
-1,
1
],
"profile": true,
"batch_size": 12,
"lr": [
0.001,
5e-06,
31.0,
0.001,
5e-06,
31.0
],
"num_epochs": 30,
"scheduler_step_size": 30,
"v1_multiscale": false,
"avg_reprojection": false,
"disable_automasking": false,
"predictive_mask": false,
"no_ssim": false,
"mypretrain": null,
"weights_init": "pretrained",
"pose_model_input": "pairs",
"pose_model_type": "separate_resnet",
"no_cuda": false,
"num_workers": 12,
"load_weights_folder": "pretrain_weights/lite-mono-8m_640_192/",
"models_to_load": [
"encoder",
"depth"
],
"log_frequency": 10,
"save_frequency": 3,
"disable_median_scaling": false,
"pred_depth_scale_factor": 1,
"ext_disp_to_eval": null,
"eval_split": "eigen",
"save_pred_disps": false,
"no_eval": false,
"eval_out_dir": null,
"post_process": false
}
Since the depth range of your image is small, so you need to change max_depth
in the configuration. The most important thing is to change the camera intrinsics in the dataset class.
Thank you for your help!! And I have other questions:
(1) What is the unit of this parameter max_depth
, centimeters or meters?
(2) How can I obtain KITTIDataset(). K
from the camera?
1) meter 2) If you collect your dataset using your own camera, then you need to calibrate your camera. There are a lot of documentations on camera calibration.
Okay, thank you very much! I'll train it again.
I am now closing this issue as there is no further update.
Hello, I imitated KITTIOdomDataset to read my own video frames, but the output images of test_simple.py are all black or yellow. What is the problem? I train the following pretrained models and get the results: