Добавлены пропсы конвейера и стереодвижки, задействованные в прогоне

assets/conveyors (274 МБ) - ленты и угловая секция NVIDIA, на которые ссылается сцена
относительным путём. Раньше исключались как перекачиваемые, но без них сцена не
композится из коробки.

cv/ - код стереодвижков, которые вызывает control_test, без весов:
* defom-stereo - рабочий бейзлайн (DEFOM vitl, вход 480, iters 24)
* crestereo - второй движок, точнее по габаритам (MAE 23.5 против 32.8 мм)
* fast-foundationstereo - проверялся, в бейзлайн не вошёл
* circular_section.py - показатель кругового сечения, перенесён в measure_plane.py:
  выравнивает облако по СОБСТВЕННЫМ главным осям и режет на пяти высотах вдоль каждой.
  Три самодельные версии (мировые оси, одно сечение) давали хуже; результаты проверки
  на эталонной геометрии - в circular_section_results.json

Веса по-прежнему не в репозитории - источники в MODELS.md. Наборы кадров прежних
прогонов (cv/flow_*, 1.26 ГБ) исключены: это выход, а не исходники.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

cv/defom-stereo/scripts/download_dav2.sh

@@ -0,0 +1,6 @@
+cd checkpoints
+wget https://huggingface.co/depth-anything/Depth-Anything-V2-Small/resolve/main/depth_anything_v2_vits.pth
+wget https://huggingface.co/depth-anything/Depth-Anything-V2-Large/resolve/main/depth_anything_v2_vitl.pth
+cd ..
+
+

cv/defom-stereo/scripts/download_models.sh

@@ -0,0 +1,10 @@
+cd checkpoints
+gdown https://drive.google.com/uc?id=1XuAM4vqzura_6NKN70hMW5lFD4TafnDL
+gdown https://drive.google.com/uc?id=1FNt-SDysG5bUOmjZ91mzH2V_TXvLCvr5
+gdown https://drive.google.com/uc?id=1qyXKO-Nxq3ndl2H0deQpo6BSvwlGKYEg
+gdown https://drive.google.com/uc?id=1Dy1eGDdtkp2GQYQRTvMwR-3eAzaRCe_k
+gdown https://drive.google.com/uc?id=1duHLtUCDNIA76m6Fqwa7hv-aBMY-P3mg
+gdown https://drive.google.com/uc?id=1xEPS7gceJSFn_IHdzebBCgQaRNGwf1aG
+cd ..
+
+

cv/defom-stereo/scripts/evaluate.sh

@@ -0,0 +1,97 @@
+#!/usr/bin/env bash
+
+# evalutate on scene flow
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vitl_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets things \
+--dinov2_encoder vitl 
+
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vits_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets things \
+--dinov2_encoder vits
+
+# evalutate on kitti12, kitti15, and eth3d
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vitl_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets kitti12 kitti15 eth3d \
+--dinov2_encoder vitl 
+
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vits_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets kitti12 kitti15 eth3d \
+--dinov2_encoder vits
+
+
+# evalutate on Middlebury; when evaluating defomstereo_vitl on Middlebury_F
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vitl_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets middlebury_F middlebury_H middlebury_Q \
+--dinov2_encoder vitl 
+
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vits_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets middlebury_F middlebury_H middlebury_Q \
+--dinov2_encoder vits
+
+# evalutate on different region.
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vitl_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets middlebury_F middlebury_H middlebury_Q \
+--indetail \
+--dinov2_encoder vitl 
+
+python evaluate_stereo.py \
+--restore_ckpt checkpoints/defomstereo_vits_sceneflow.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets middlebury_F middlebury_H middlebury_Q \
+--indetail \
+--dinov2_encoder vits
+

cv/defom-stereo/scripts/make_submission.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+
+# make submission to stereo benchmarks
+
+# make submission for  kitti12 and kitti15
+python make_submission.py \
+--restore_ckpt checkpoints/defomstereo_vitl_kitti.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets kitti12 kitti15 \
+--dinov2_encoder vitl 
+
+# make submission for eth3d
+python make_submission.py \
+--restore_ckpt checkpoints/defomstereo_vitl_eth3d.pth  \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets eth3d \
+--dinov2_encoder vitl
+
+# make submission for middlebury
+python make_submission.py \
+--restore_ckpt checkpoints/defomstereo_vitl_middlebury.pth  \
+--method_name DEFOM-Stereo \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets middlebury_F \
+--dinov2_encoder vitl
+
+# make submission for kitti15, middlebury and eth3d using the RVC model
+python make_submission.py \
+--restore_ckpt checkpoints/defomstereo_vits_rvc.pth  \
+--method_name DEFOM-Stereo_RVC \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--datasets kitti15 middlebury_F eth3d \
+--dinov2_encoder vits

cv/defom-stereo/scripts/train_eth3d.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+
+# trained on 4 x 24GB 3090/4090 GPUs
+
+CHECKPOINT_DIR=checkpoints/defomstereo_vitl_eth3d_pretrain && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=4 --master_port=9994 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 \
+--name defomstereo_vitl_eth3d_pretrain \
+--batch_size 8  \
+--num_workers 8  \
+--train_datasets tartan_air sceneflow sintel_stereo eth3d instereo2k crestereo \
+--train_folds 1 1 50 1000 100 2 \
+--num_steps 300000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vitl \
+--image_size 384 512 \
+--resume_ckpt checkpoints/defomstereo_vitl_sceneflow.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log && \
+
+CHECKPOINT_DIR=checkpoints/defomstereo_vitl_eth3d && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=4 --master_port=9993 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 \
+--name defomstereo_vitl_eth3d \
+--batch_size 8  \
+--num_workers 8  \
+--train_datasets eth3d instereo2k crestereo \
+--train_folds 1000 10 1 \
+--num_steps 90000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vitl \
+--image_size 384 512 \
+--resume_ckpt checkpoints/defomstereo_vitl_eth3d_pretrain.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log 

cv/defom-stereo/scripts/train_kitti.sh

@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+
+# trained on 8 x 24GB 3090/4090 GPUs
+
+CHECKPOINT_DIR=checkpoints/defomstereo_vitl_kitti && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=8 --master_port=9992 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 4 5 6 7 \
+--name defomstereo_vitl_kitti \
+--batch_size 8  \
+--num_workers 4  \
+--train_datasets kitti12 kitti15 vkitti2 \
+--train_folds 50 50 1 \
+--num_steps 50000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vitl \
+--image_size 352 1216 \
+--resume_ckpt checkpoints/defomstereo_vitl_sceneflow.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
+
+

cv/defom-stereo/scripts/train_middlebury.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+
+# trained on 4 x 24GB 3090/4090 GPUs
+CHECKPOINT_DIR=checkpoints/defomstereo_vitl_middlebury_pretrain && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=4 --master_port=9993 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 \
+--name defomstereo_vitl_middlebury_pretrain \
+--batch_size 8  \
+--num_workers 8  \
+--train_datasets tartan_air sceneflow falling_things instereo2k carla_highres crestereo middlebury_2014 middlebury_2021 middlebury_H \
+--train_folds 1 1 1 50 50 1 200 200 200 \
+--num_steps 200000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vitl \
+--image_size 384 512 \
+--resume_ckpt checkpoints/defomstereo_vitl_sceneflow.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log && \
+
+# trained on 8 x 24GB 3090/4090 GPUs
+CHECKPOINT_DIR=checkpoints/defomstereo_vitl_middlebury && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=8 --master_port=9993 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 4 5 6 7 \
+--name defomstereo_vitl_middlebury \
+--batch_size 8  \
+--num_workers 4  \
+--train_datasets crestereo instereo2k carla_highres middlebury_2014 middlebury_2021 middlebury_H middlebury_F falling_things \
+--train_folds 1 50 50 200 200 200 200 5  \
+--num_steps 100000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vitl \
+--image_size 512 768 \
+--resume_ckpt checkpoints/defomstereo_vitl_middlebury_pretrain.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log 

cv/defom-stereo/scripts/train_rvc.sh

@@ -0,0 +1,79 @@
+#!/usr/bin/env bash
+
+# trained on 4 x 24GB 3090/4090 GPUs
+CHECKPOINT_DIR=checkpoints/defomstereo_vits_rvc_pretrain && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=4 --master_port=9995 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 \
+--name defomstereo_vits_rvc_pretrain \
+--batch_size 8  \
+--num_workers 8  \
+--train_datasets tartan_air sceneflow irs 3dkenburns crestereo falling_things sintel_stereo vkitti2 carla_highres \
+--train_folds 1 1 1 1 1 1 3 3 80 \
+--num_steps 200000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vits \
+--image_size 384 768 \
+--resume_ckpt checkpoints/defomstereo_vits_sceneflow.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log && \
+
+# trained on 4 x 24GB 3090/4090 GPUs
+CHECKPOINT_DIR=checkpoints/defomstereo_vits_rvc_pretrain2 && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=4 --master_port=9996 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 \
+--name defomstereo_vits_rvc_pretrain2 \
+--batch_size 8  \
+--num_workers 8  \
+--train_datasets tartan_air irs 3dkenburns crestereo vkitti2 carla_highres kitti12 kitti15 middlebury_2005 middlebury_2006 middlebury_2014 middlebury_2021 middlebury_Q middlebury_H eth3d instereo2k booster \
+--train_folds 1 1 1 1 3 30 100 100 200 200 200 200 200 200 1000 20 10 \
+--num_steps 100000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vits \
+--image_size 384 768 \
+--resume_ckpt checkpoints/defomstereo_vits_rvc_pretrain.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log && \
+
+# trained on 4 x 24GB 3090/4090 GPUs
+CHECKPOINT_DIR=checkpoints/defomstereo_vits_rvc && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=4 --master_port=9997 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 \
+--name defomstereo_vits_rvc \
+--batch_size 8  \
+--num_workers 8  \
+--train_datasets tartan_air irs 3dkenburns crestereo vkitti2 carla_highres kitti12 kitti15 middlebury_2005 middlebury_2006 middlebury_2014 middlebury_2021 middlebury_Q middlebury_H eth3d instereo2k booster \
+--train_folds 1 1 1 1 3 30 2500 2500 200 200 200 200 200 200 1000 20 10 \
+--num_steps 20000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vits \
+--image_size 384 768 \
+--resume_ckpt checkpoints/defomstereo_vits_rvc_pretrain2.pth \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log 

cv/defom-stereo/scripts/train_sceneflow_vitl.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+
+# trained on 4 x 24GB 3090/4090 GPUs
+
+CHECKPOINT_DIR=checkpoints/defomstereo_vitl_sceneflow && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=4 --master_port=9991 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 2 3 \
+--name defomstereo_vitl_sceneflow \
+--batch_size 8  \
+--num_workers 8  \
+--train_datasets sceneflow \
+--train_folds 1 \
+--num_steps 200000 \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vitl \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
+
+

cv/defom-stereo/scripts/train_sceneflow_vits.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+
+# trained on 2 x 24GB 3090/4090 GPUs
+
+CHECKPOINT_DIR=checkpoints/defomstereo_vits_sceneflow && \
+mkdir -p ${CHECKPOINT_DIR} && \
+python -m torch.distributed.launch --nproc_per_node=2 --master_port=9990 train_stereo.py \
+--distributed \
+--launcher pytorch \
+--gpu_ids 0 1 \
+--name defomstereo_vits_sceneflow \
+--batch_size 8  \
+--num_workers 16  \
+--train_datasets sceneflow \
+--train_folds 1 \
+--num_steps 200000 \
+--mixed_precision \
+--n_downsample 2 \
+--train_iters 18 \
+--scale_iters 8 \
+--idepth_scale 0.5 \
+--corr_levels 2 \
+--corr_radius 4 \
+--scale_list 0.125 0.25 0.5 0.75 1.0 1.25 1.5 2.0 \
+--scale_corr_radius 2 \
+--dinov2_encoder vits \
+2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
+
+