A compact research repository for referring video object segmentation (RVOS), combining category grounding, SAM3 candidate recall, and candidate-aware reranking.
referring expression + video frames
-> category grounding
-> SAM3 candidate tube recall
-> candidate-aware visual scoring
-> mask selection and evaluation
| Path | Contents |
|---|---|
Sa2VA-main/category_grounding_mvp/ | Category grounding, data construction, and inference |
Sa2VA-main/crf_sam3_attn/ | Candidate reranker, training, and cache metadata |
Sa2VA-main/SAM3-DMS/ | Candidate recall implementation |
Sa2VA-main/configs/ | Dataset and inference configurations |
Sa2VA-main/tools/eval/ | Evaluation tools |
Sa2VA-main/crf_sam3_*/ | Evaluation summaries and selected supporting records |
Sa2VA/data/video_datas/ | Dataset layout documentation |
论文代理/ | Paper figures and analysis |
See PAPER_EVIDENCE_MAP.md for source and result entry points.
Run from Sa2VA-main/ after configuring the data paths, base models, learned weights, and Python environment.
Three-dataset reranker training:
python -u crf_sam3_attn/train.py --config crf_sam3_attn/config_joint_train_train.yaml
ReVOS reranker training:
bash crf_sam3_attn/run_revos_train_warmstart_3domain_api_categories_recall24_keep7_20260805.sh
The ReVOS inference configuration is configs/crf_sam3_config_revos_valid_revos_query_ranker_20260806.yaml; evaluation is implemented in tools/eval/eval_revos.py.
| Dataset / split | J | F | J&F | Evaluation file under Sa2VA-main/ |
|---|---|---|---|---|
| MeViS valid_u | 68.83 | 74.07 | 71.45 | crf_sam3_mevis_valid_u_joint_train_poswin_mergedprompt/evaluation_results.json |
| DAVIS17 valid | 74.06 | 78.73 | 76.39 | crf_sam3_davis17_valid_query_grounding_semantic_expand_topk5_final_20260703/evaluation_results_davis_png.json |
| Ref-Youtube-VOS valid (overlay) | 65.23 | 67.72 | 66.47 | crf_sam3_refytros_valid_query_grounding_semantic_expand_topk5_failures_only_20260702/overlay_evaluation_results_refytros_png.json |
| ReVOS valid | 57.20 | 62.28 | 59.74 | crf_sam3_revos_valid_revos_query_ranker_20260806/revos_valid_official.json |
The first three rows use expression-level metrics. The Ref-Youtube-VOS overlay combines 666 base expressions with 168 patched expressions. ReVOS uses the official overall aggregation; its audit covers 5,822 expressions with zero invalid predictions. Each result is tied to its own run configuration and evaluation protocol.
See dataset documentation for splits, annotation layouts, metadata counts, and local sample coverage.
Git retains source code, configurations, dataset layout documentation, sampled cache metadata, evaluation summaries, and paper materials. Raw media, mask dictionaries, checkpoints, tokenizer payloads, training logs, generated training sets, and full prediction payloads are excluded.
Datasets include MeViS, DAVIS17, Ref-Youtube-VOS, and ReVOS; revos_prepared is a derived ReVOS training layout. sample_inventory.json describes local samples, which do not replace full evaluation datasets. Full reproduction requires the original datasets and model weights.
A compact research repository for referring video object segmentation (RVOS), combining category grounding, SAM3 candidate recall, and candidate-aware reranking.
referring expression + video frames
-> category grounding
-> SAM3 candidate tube recall
-> candidate-aware visual scoring
-> mask selection and evaluation
| Path | Contents |
|---|---|
Sa2VA-main/category_grounding_mvp/ | Category grounding, data construction, and inference |
Sa2VA-main/crf_sam3_attn/ | Candidate reranker, training, and cache metadata |
Sa2VA-main/SAM3-DMS/ | Candidate recall implementation |
Sa2VA-main/configs/ | Dataset and inference configurations |
Sa2VA-main/tools/eval/ | Evaluation tools |
Sa2VA-main/crf_sam3_*/ | Evaluation summaries and selected supporting records |
Sa2VA/data/video_datas/ | Dataset layout documentation |
论文代理/ | Paper figures and analysis |
See PAPER_EVIDENCE_MAP.md for source and result entry points.
Run from Sa2VA-main/ after configuring the data paths, base models, learned weights, and Python environment.
Three-dataset reranker training:
python -u crf_sam3_attn/train.py --config crf_sam3_attn/config_joint_train_train.yaml
ReVOS reranker training:
bash crf_sam3_attn/run_revos_train_warmstart_3domain_api_categories_recall24_keep7_20260805.sh
The ReVOS inference configuration is configs/crf_sam3_config_revos_valid_revos_query_ranker_20260806.yaml; evaluation is implemented in tools/eval/eval_revos.py.
| Dataset / split | J | F | J&F | Evaluation file under Sa2VA-main/ |
|---|---|---|---|---|
| MeViS valid_u | 68.83 | 74.07 | 71.45 | crf_sam3_mevis_valid_u_joint_train_poswin_mergedprompt/evaluation_results.json |
| DAVIS17 valid | 74.06 | 78.73 | 76.39 | crf_sam3_davis17_valid_query_grounding_semantic_expand_topk5_final_20260703/evaluation_results_davis_png.json |
| Ref-Youtube-VOS valid (overlay) | 65.23 | 67.72 | 66.47 | crf_sam3_refytros_valid_query_grounding_semantic_expand_topk5_failures_only_20260702/overlay_evaluation_results_refytros_png.json |
| ReVOS valid | 57.20 | 62.28 | 59.74 | crf_sam3_revos_valid_revos_query_ranker_20260806/revos_valid_official.json |
The first three rows use expression-level metrics. The Ref-Youtube-VOS overlay combines 666 base expressions with 168 patched expressions. ReVOS uses the official overall aggregation; its audit covers 5,822 expressions with zero invalid predictions. Each result is tied to its own run configuration and evaluation protocol.
See dataset documentation for splits, annotation layouts, metadata counts, and local sample coverage.
Git retains source code, configurations, dataset layout documentation, sampled cache metadata, evaluation summaries, and paper materials. Raw media, mask dictionaries, checkpoints, tokenizer payloads, training logs, generated training sets, and full prediction payloads are excluded.
Datasets include MeViS, DAVIS17, Ref-Youtube-VOS, and ReVOS; revos_prepared is a derived ReVOS training layout. sample_inventory.json describes local samples, which do not replace full evaluation datasets. Full reproduction requires the original datasets and model weights.