Can not detect custom object using inference.py

[patsyuk03]

Hello.
I have generated data using this command:
python single_video_pybullet.py --nb_frames 10000 --scale 0.001 --path_single_obj ~/Deep_Object_Pose/scripts/nvisii_data_gen/models/Gear/google_16k/gear.obj --nb_distractors 0 --nb_object 10 --outf gear1/

And trained the model for 60 epoch on 9800 generated images:
python -m torch.distributed.launch --nproc_per_node=1 train.py --network dope --epochs 60 --batchsize 2 --outf tmp_gear1/ --data ../nvisii_data_gen/output/gear1/

When I run inference on the rest 200 generated images. The belief maps seems good, but there are no objects detected.

Here is the inference config:

> topic_camera: "/dope/webcam/image_raw"
> topic_camera_info: "/dope/webcam/camera_info"
> topic_publishing: "dope"
> input_is_rectified: True   # Whether the input image is rectified (strongly suggested!)
> downscale_height: 400      # if the input image is larger than this, scale it down to this pixel height
> 
> # Comment any of these lines to prevent detection / pose estimation of that object
> weights: {
>     #'obj':"tmp/net_epoch_99.pth"
>     'obj':"tmp_gear1/net_epoch_60.pth"
> }
> 
> # Type of neural network architecture
> architectures: {
>     'obj':"dope",
> }
> 
> 
> # Cuboid dimension in cm x,y,z
> dimensions: {
>     'obj':[11.9541015625, 3.00, 11.869906616210938]
> 
> }
> 
> class_ids: {
>     "obj": 1
> }
> 
> draw_colors: {
>     "obj": [13, 255, 128],  # green
> }
> 
> # optional: provide a transform that is applied to the pose returned by DOPE
> model_transforms: {
> #    "cracker": [[ 0,  0,  1,  0],
> #                [ 0, -1,  0,  0],
> #                [ 1,  0,  0,  0],
> #                [ 0,  0,  0,  1]]
> }
> 
> # optional: if you provide a mesh of the object here, a mesh marker will be
> # published for visualization in RViz
> # You can use the nvdu_ycb tool to download the meshes: https://github.com/NVIDIA/Dataset_Utilities#nvdu_ycb
> meshes: {
> #    "cracker": "file://path/to/Dataset_Utilities/nvdu/data/ycb/aligned_cm/003_cracker_box/google_16k/textured.obj",
> #    "gelatin": "file://path/to/Dataset_Utilities/nvdu/data/ycb/aligned_cm/009_gelatin_box/google_16k/textured.obj",
> #    "meat":    "file://path/to/Dataset_Utilities/nvdu/data/ycb/aligned_cm/010_potted_meat_can/google_16k/textured.obj",
> #    "mustard": "file://path/to/Dataset_Utilities/nvdu/data/ycb/aligned_cm/006_mustard_bottle/google_16k/textured.obj",
> #    "soup":    "file://path/to/Dataset_Utilities/nvdu/data/ycb/aligned_cm/005_tomato_soup_can/google_16k/textured.obj",
> #    "sugar":   "file://path/to/Dataset_Utilities/nvdu/data/ycb/aligned_cm/004_sugar_box/google_16k/textured.obj",
> #    "bleach":  "file://path/to/Dataset_Utilities/nvdu/data/ycb/aligned_cm/021_bleach_cleanser/google_16k/textured.obj",
> }
> 
> # optional: If the specified meshes are not in meters, provide a scale here (e.g. if the mesh is in centimeters, scale should be 0.01). default scale: 1.0.
> mesh_scales: {
>     "obj": 0.01
> }
> 
> # Config params for DOPE
> thresh_angle: 0.5
> thresh_map: 0.0001
> sigma: 3
> thresh_points: 0.1

Is there something that I can do to fix this?

[TontonTremblay]

wow this is such a beautiful example of training on a symmetrical object. You can fix this by retraining using the script to annotate the symmetries.

https://github.com/NVlabs/Deep_Object_Pose/tree/master/data_generation#handling-objects-with-symmetries

[patsyuk03]

Thank you for the quick answer.

As I understand, it is similar to your example of a Hex screw object with rotational symmetry. However, my object, due to the hexagon in the center, is not entirely symmetrical rotationally, but I can see the centerline where it can be mirrored.

What would be the right way to define the symmetry in this case? Will it be possible for a model to distinguish between such small offsets of hexagon corners, or is the only option to just ignore it and define it as symmetrical rotationally?

[TontonTremblay]

There is an axe for each hexagon corner.